Intelligent design and analysis method and system of anti-bending photonic crystal fiber

Abstract

The invention relates to an intelligent design and analysis method and system of an anti-bending photonic crystal fiber. The intelligent design method comprises the following steps: generating a first-dimensional structure parameter of a bending-resistant photonic crystal fiber and a second-dimensional fiber bending-resistant performance parameter corresponding to the first-dimensional structure parameter, wherein the dimension of the second-dimensional fiber bending-resistant performance parameter is greater than that of the first-dimensional structure parameter; based on the first-dimensional structural parameters and the corresponding second-dimensional optical fiber bending-resistant performance parameters, constructing a bending-resistant photonic crystal optical fiber intelligent reverse design model for predicting the first-dimensional structural parameters; calculating a structure parameter fitting coefficient corresponding to the bending-resistant photonic crystal fiber intelligent reverse design model; and constructing a bending-resistant photonic crystal fiber intelligent reverse design optimization model which is not lower than the structure parameter fitting coefficient and a corresponding minimum-dimension fiber bending-resistant performance parameter in the second-dimension fiber bending-resistant performance parameters. According to the embodiment of the invention, intelligent design and analysis of the bending-resistant photonic crystal fiber can be realized.

Description

Technical Field

[0001] The present disclosure relates to the technical field of anti-bending photonic crystal fiber design, and particularly to an intelligent design, analysis method and system for anti-bending photonic crystal fibers. Background Art

[0002] Photonic Crystal Fiber (PCF) has shown broad application prospects in the fields of optical communication, sensing, nonlinear optics, etc. due to its unique waveguide structure and excellent optical properties. However, traditional photonic crystal fibers are prone to large bending losses under bending conditions, which greatly limits their applications in certain specific scenarios. For example, in medical imaging devices, optical fibers need to flexibly penetrate deep into the human body for light conduction to obtain image information, and frequent bending operations are normal. However, due to large bending losses of traditional PCF, the intensity of the optical signal transmitted to the detection end is greatly weakened, and the imaging quality is greatly reduced, unable to meet the high-precision requirements of clinical diagnosis. Another example is in the aerospace field, where the internal space of the aircraft is compact, and the wiring requires the optical fiber to be bent multiple times to adapt to the complex spatial structure. The bending losses of traditional PCF will make the signal transmission unstable, affecting the normal operation of the key systems of the aircraft. To overcome this challenge, the existing technology mainly improves the anti-bending performance of photonic crystal fibers by optimizing the waveguide structure. These optimization means include adjusting the size, shape and arrangement of air holes, as well as changing the refractive index distribution of the core and cladding, etc.

[0003] The design flexibility of parameters such as the air hole diameter and air hole spacing brings more possibilities to the design of anti-bending photonic crystal fibers. It effectively improves the performance under bending conditions through the careful design of parts such as the air hole structure, core and cladding. However, this high flexibility of the structural design makes the anti-bending performance not simply change with the change of the fiber structure parameters. When designing anti-bending photonic crystal fibers, there are many variable parameters, such as the air hole spacing, different air hole diameters, the selection of fiber materials, the total fiber diameter, etc., and many properties that are interrelated and interact with each other, such as bending loss, mode characteristics, dispersion, etc., must be considered. This leads to an exponential increase in the design difficulty with the increase of their quantity. Currently, the design of anti-bending photonic crystal fibers mainly relies on experimental-based methods. Designers need to frequently use finite element software such as COMSOL Multiphysics for simulation calculations and continuously adjust the design parameters. In this process, designers not only spend a lot of time and energy interacting with the software calculation process, but also cause a great burden on the computer calculation resources.

[0004] With the rapid development of artificial intelligence technology, machine learning (ML) technology has been widely used in the field of photonic crystal fiber (PCF) design. It can handle complex nonlinear relationships, and by learning a large amount of PCF structural parameters and performance data, it can quickly establish a mapping relationship between the two, and achieve efficient prediction of PCF performance and optimized structural design. However, machine learning has also exposed serious defects in PCF design applications. The internal decision-making models constructed by these machine learning algorithms have not yet been fully understood. In practical applications, although the model can give a prediction result, it is difficult to explain why it makes such a prediction instead of other predictions. This phenomenon is not accidental, but due to the interaction of a large number of internal variables during the training process, forming intricate relationships. These relationships are hidden deep in the model and are difficult for researchers to interpret and analyze.

[0005] This lack of explainability has become a key constraint for traditional and ML-based design strategies. In scientific research, it is difficult for researchers to deeply understand the decision-making basis of the model, and it is impossible to optimize and improve the design from the physical principle level. In the field of engineering applications, the lack of explainability makes it difficult to fully verify the reliability and stability of the design scheme, increasing the risk of practical application. Therefore, how to interpret the structural parameters in the design of anti-bending PCF based on the second-dimensional optical fiber anti-bending performance parameters and machine learning algorithms, and then determine the corresponding structural parameters of the anti-bending PCF design, has become an important issue that needs to be solved in this field. Summary of the invention

[0006] The present disclosure proposes a technical solution for an intelligent design, analysis method and system of bend-resistant photonic crystal optical fibers.

[0007] According to one aspect of the present disclosure, there is provided an intelligent design method for a bend-resistant photonic crystal fiber, which is applied to a reverse prediction task, comprising:

[0008] Generate first-dimensional structural parameters of a bend-resistant photonic crystal fiber and its corresponding second-dimensional fiber bend resistance performance parameters of a dimension greater than the first-dimensional structural parameters;

[0009] Based on the first-dimensional structural parameters and the corresponding second-dimensional optical fiber bending resistance performance parameters, constructing a bending-resistant photonic crystal fiber intelligent reverse design model for predicting the first-dimensional structural parameters;

[0010] Calculating the structural parameter fitting coefficients corresponding to the bending-resistant photonic crystal fiber intelligent reverse design model;

[0011] Construct an anti-bending photonic crystal fiber intelligent reverse design optimization model not lower than the fitting coefficient of the structural parameters and the minimum fiber anti-bending performance parameter among the corresponding second-dimensional fiber anti-bending performance parameters.

[0012] Preferably, the method for generating the first-dimensional structural parameters of the anti-bending photonic crystal fiber and the corresponding second-dimensional fiber anti-bending performance parameters of dimensions greater than the first-dimensional structural parameters includes: using physical field simulation software, under the set working wavelength and the first-dimensional structural parameters corresponding to the anti-bending photonic crystal fiber, adjusting the bending radius corresponding to the photonic crystal fiber to generate the first-dimensional structural parameters of the anti-bending photonic crystal fiber and the corresponding second-dimensional fiber anti-bending performance parameters of dimensions greater than the first-dimensional structural parameters.

[0013] Preferably, the method for adjusting the bending radius corresponding to the photonic crystal fiber under the set working wavelength and the first-dimensional structural parameters corresponding to the anti-bending photonic crystal fiber to generate the first-dimensional structural parameters of the anti-bending photonic crystal fiber and the corresponding second-dimensional fiber anti-bending performance parameters includes:

[0014] Step S201: Configure the first-dimensional structural parameters as the first diameter of the air holes corresponding to the first size in the anti-bending photonic crystal fiber, the second diameter of the air holes corresponding to the second size smaller than the first size, and the air hole spacing between the air holes;

[0015] Step S202: Obtain the set range and set step length corresponding to the first diameter, the second diameter, the air hole spacing, and the bending radius;

[0016] Step S203: Configure any one of the structural parameters among the first diameter, the second diameter, and the air hole spacing as a variable parameter, and configure the structural parameters other than the variable parameter as constant parameters;

[0017] Step S204: Starting from any endpoint of the set range corresponding to the variable parameter, under the set working wavelength, adjust the bending radius corresponding to the photonic crystal fiber, and determine the parameter change corresponding to the variable parameter according to the corresponding set step;

[0018] Step S205: Determine the corresponding second-dimensional fiber anti-bending performance parameter under the parameter change corresponding to the variable parameter and the constant parameters;

[0019] Loop or repeat the above steps S203 - S205 until the first diameter, the second diameter, and the air hole spacing are configured as variable parameters, thereby generating the first-dimensional structural parameters of the anti-bending photonic crystal fiber and the corresponding second-dimensional fiber anti-bending performance parameters.

[0020] Preferably, before constructing an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters, logarithmic processing is performed on the first bending loss of the high-order mode and the second bending loss of the fundamental mode in the second-dimensional fiber bend-resistant performance parameters to obtain the logarithmic first bending loss and the logarithmic second bending loss; and normalization processing is performed on the logarithmic first bending loss, the logarithmic second bending loss, and the effective mode area in the second-dimensional fiber bend-resistant performance parameters.

[0021] Preferably, the second-dimensional fiber bend-resistant performance parameters include one or several of the first bending loss of the high-order mode, the second bending loss of the fundamental mode, and the effective mode area corresponding to the bend-resistant photonic crystal fiber under the first-dimensional structural parameters or different structural parameters.

[0022] Preferably, the method for constructing an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters includes: determining a training set based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters; and using the training set to train a preset machine learning algorithm or a preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber to construct an intelligent inverse design model of the bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters.

[0023] Preferably, the method for calculating the structural parameter fitting coefficient corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber includes: determining a test set based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters; configuring the performance index corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber as the structural parameter fitting coefficient; and using the test set to calculate the structural parameter fitting coefficient corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber.

[0024] Preferably, the method for constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter among the corresponding second-dimensional fiber bend-resistant performance parameters includes: using a global interpretability algorithm to interpret the second-dimensional fiber bend-resistant performance parameters, each fiber bend-resistant performance parameter of each dimensional fiber bend-resistant performance parameter in the second-dimensional fiber bend-resistant performance parameters; sorting the second-dimensional fiber bend-resistant performance parameters based on the contribution degree corresponding to each fiber bend-resistant performance parameter to obtain a sorted second-dimensional fiber bend-resistant performance parameter; selecting corresponding updated-dimensional fiber bend-resistant performance parameters from the sorted second-dimensional fiber bend-resistant performance parameters according to the contribution degree corresponding to each fiber bend-resistant performance parameter and a set contribution degree; and constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter among the corresponding second-dimensional fiber bend-resistant performance parameters based on the updated-dimensional fiber bend-resistant performance parameters, the first-dimensional structural parameters, and the corresponding optimized preset machine learning algorithm or preset network.

[0025] Preferably, the method for constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter among the corresponding second-dimensional fiber bend-resistant performance parameters based on the updated-dimensional fiber bend-resistant performance parameters, the first-dimensional structural parameters, and the corresponding optimized preset machine learning algorithm or preset network includes:

[0026] Step 301: Optimize the preset machine learning algorithm or preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber using the updated-dimensional fiber bend-resistant performance parameters; Step 302: Train the optimized preset machine learning algorithm or preset network using the updated-dimensional fiber bend-resistant performance parameters and their corresponding first-dimensional structural parameters; Step 303: Calculate the optimized structural parameter fitting coefficient of the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber corresponding to the optimized preset machine learning algorithm or preset network; Step 304: Compare the optimized structural parameter fitting coefficient with the structural parameter fitting coefficient corresponding to the second-dimensional fiber bend-resistant performance parameters; Step 304: If the optimized structural parameter fitting coefficient is greater than or equal to the structural parameter fitting coefficient corresponding to the second-dimensional fiber bend-resistant performance parameters, obtain a set deletion dimension; delete the fiber bend-resistant performance parameters corresponding to the set deletion dimension that are ranked lower from the updated-dimensional fiber bend-resistant performance parameters to obtain the latest updated-dimensional fiber bend-resistant performance parameters; where the set deletion dimension is greater than or equal to 1 and is a positive integer;

[0027] Based on the anti-bending performance parameters of the latest updated dimension fiber; repeat steps 301 to 304 until an intelligent reverse design optimization model of the anti-bending photonic crystal fiber not lower than the anti-bending performance parameter of the minimum dimension fiber corresponding to the structural parameter fitting coefficient and the corresponding second-dimension fiber anti-bending performance parameter are constructed;

[0028] Step 305: If the optimized structural parameter fitting coefficient is less than the structural parameter fitting coefficient corresponding to the second-dimension fiber anti-bending performance parameter, obtain a set of added dimensions; add the fiber anti-bending performance parameters corresponding to the set of deleted dimensions with the highest ranking from the updated dimension fiber anti-bending performance parameters to obtain the latest updated dimension fiber anti-bending performance parameters; wherein, the set of added dimensions is greater than or equal to 1 and is a positive integer;

[0029] Based on the latest updated dimension fiber anti-bending performance parameters, repeat steps 301 to 303 and step 305 until an intelligent reverse design optimization model of the anti-bending photonic crystal fiber not lower than the anti-bending performance parameter of the minimum dimension fiber corresponding to the structural parameter fitting coefficient and the corresponding second-dimension fiber anti-bending performance parameter are constructed.

[0030] Preferably, the method for optimizing the preset machine learning algorithm or preset network corresponding to the intelligent design of the anti-bending photonic crystal fiber by using the updated dimension fiber anti-bending performance parameters includes:

[0031] When selecting the corresponding updated-dimensional fiber bending resistance performance parameter from the second-ranked fiber bending resistance performance parameters or deleting the fiber bending resistance performance parameter corresponding to the set deletion dimension with a lower ranking from the updated-dimensional fiber bending resistance performance parameter to obtain the latest updated-dimensional fiber bending resistance performance parameter, adjust the number of neurons in the input layer corresponding to the preset machine learning algorithm or preset network for the intelligent design of the bending-resistant photonic crystal fiber using the updated-dimensional fiber bending resistance performance parameter or the latest updated-dimensional fiber bending resistance performance parameter; according to the set layer step size, reduce the number of layers and the number of neurons in each layer of the middle hidden layer of the preset machine learning algorithm or preset network for the intelligent design of the bending-resistant photonic crystal fiber to obtain an optimized preset machine learning algorithm or preset network; wherein, the number of neurons in the input layer is the same as the dimension corresponding to the updated-dimensional fiber bending resistance performance parameter or the latest updated-dimensional fiber bending resistance performance parameter; or, when adding the fiber bending resistance performance parameter corresponding to the set deletion dimension with a higher ranking from the updated-dimensional fiber bending resistance performance parameter to obtain the latest updated-dimensional fiber bending resistance performance parameter, adjust the number of neurons in the input layer corresponding to the preset machine learning algorithm or preset network for the intelligent design of the bending-resistant photonic crystal fiber using the updated-dimensional fiber bending resistance performance parameter; according to the set layer step size, increase the number of layers and the number of neurons in each layer of the middle hidden layer of the preset machine learning algorithm or preset network for the intelligent design of the bending-resistant photonic crystal fiber to obtain an optimized preset machine learning algorithm or preset network; wherein, the number of neurons in the input layer is the same as the dimension corresponding to the latest updated-dimensional fiber bending resistance performance parameter.

[0032] Preferably, the middle hidden layer of the optimized preset machine learning algorithm or preset network includes: a first middle hidden layer connected to the input layer, a plurality of second middle hidden layers, and a last middle hidden layer connected to the output layer of the preset machine learning algorithm or preset network for the intelligent design of the bending-resistant photonic crystal fiber; wherein, a first activation function is configured after each neuron in the first middle hidden layer; a regularization function is configured after each neuron in each second middle hidden layer among the plurality of second middle hidden layers; a second activation function is configured after each neuron in the last middle hidden layer; wherein, the input of the first second middle hidden layer among the plurality of second middle hidden layers is connected to the first middle hidden layer, and the output of the first second middle hidden layer is connected to the input of the second second middle hidden layer; and so on, the output layer of the last second middle hidden layer is connected to the input of the last middle hidden layer.

[0033] Preferably, the first activation function is configured as a ReLU activation function; the second activation function is configured as a tanh activation function.

[0034] Preferably, the method for interpreting each fiber bending resistance performance parameter of each dimension of the second - dimension fiber bending resistance performance parameters by using the global interpretability algorithm includes:

[0035] Input the second - dimension fiber bending resistance performance parameters in the anti - bending photonic crystal fiber intelligent inverse design model and the training dataset for training the corresponding preset machine learning algorithm or preset network of the anti - bending photonic crystal fiber intelligent design into the kernel interpreter to construct an interpreter;

[0036] Based on the interpreter, determine the marginal contribution degree of each fiber bending resistance performance parameter of each dimension of the second - dimension fiber bending resistance performance parameters to predicting the first - dimension structural parameters by the anti - bending photonic crystal fiber intelligent inverse design model;

[0037] Perform weighted averaging on the multiple marginal contribution degrees corresponding to all dimensions of fiber bending resistance performance parameters to obtain the contribution value corresponding to each fiber bending resistance performance parameter of each dimension of the second - dimension fiber bending resistance performance parameters.

[0038] According to one aspect of the present disclosure, there is provided an intelligent design method for an anti - bending photonic crystal fiber, which is applied to a forward prediction task and includes: the intelligent design method for an anti - bending photonic crystal fiber applied to a reverse prediction task as described above; and,

[0039] Based on the first - dimension structural parameters and their corresponding second - dimension fiber bending resistance performance parameters, use the corresponding preset machine learning algorithm or preset network of the anti - bending photonic crystal fiber intelligent design to respectively construct multiple anti - bending photonic crystal fiber intelligent forward design models for predicting the second - dimension fiber bending resistance performance parameters;

[0040] Respectively calculate the multiple fiber bending resistance performance fitting coefficients corresponding to the multiple anti - bending photonic crystal fiber intelligent forward design models;

[0041] Based on the multiple fiber bending resistance performance fitting coefficients, determine the anti - bending photonic crystal fiber intelligent optimal forward design model.

[0042] According to one aspect of the present disclosure, there is provided a method for intelligent analysis of an anti - bending photonic crystal fiber, which includes: applying the intelligent design method as described above to construct an anti - bending photonic crystal fiber intelligent inverse design optimization model; and,

[0043] Using the locality interpretability algorithm, configure each fiber bending resistance performance parameter corresponding to the smallest - dimension fiber bending resistance performance parameter in the second - dimension fiber bending resistance performance parameters in the test dataset for training the anti - bending photonic crystal fiber intelligent inverse design optimization model as a specific sample;

[0044] Using the interpretation example function, adjust the eigenvalue corresponding to the specific sample to generate a corresponding perturbed sample;

[0045] Based on the intelligent inverse design optimization model of the bending-resistant photonic crystal fiber, the specific sample, the perturbed sample, and the local interpretability algorithm, determine the contribution degree of the local interpretability algorithm to the first-dimensional structural parameter;

[0046] If the contribution degree of the local interpretability algorithm to the first-dimensional structural parameter is basically the same as that determined by the global interpretability algorithm, the minimum-dimensional fiber bending-resistant performance parameter corresponding to the intelligent inverse design optimization model of the bending-resistant photonic crystal fiber is optimal.

[0047] Preferably, the method for determining the contribution degree of the local interpretability algorithm to the first-dimensional structural parameter based on the intelligent inverse design optimization model of the bending-resistant photonic crystal fiber, the specific sample, the perturbed sample, and the local interpretability algorithm includes:

[0048] Based on the intelligent inverse design optimization model of the bending-resistant photonic crystal fiber respectively, perform the first-dimensional structural parameter prediction on the specific sample and the perturbed sample to obtain a first predicted structural parameter and a second predicted structural parameter;

[0049] Calculate the similarity between the first predicted structural parameter and the second predicted structural parameter, and configure the similarity as the weight of the perturbed sample;

[0050] Based on the perturbed sample configured with the weight, train the local interpretability algorithm to obtain the contribution degree of each-dimensional fiber bending-resistant performance parameter corresponding to the minimum-dimensional fiber bending-resistant performance parameter to the first-dimensional structural parameter.

[0051] According to one aspect of the present disclosure, there is provided an intelligent design system for a bending-resistant photonic crystal fiber, which is applied to an inverse prediction task and includes:

[0052] A first generation unit for generating a first-dimensional structural parameter of a bending-resistant photonic crystal fiber and a second-dimensional fiber bending-resistant performance parameter greater than the first-dimensional structural parameter corresponding thereto;

[0053] A first construction unit for constructing an intelligent inverse design model of a bending-resistant photonic crystal fiber for predicting the first-dimensional structural parameter based on the first-dimensional structural parameter and the second-dimensional fiber bending-resistant performance parameter corresponding thereto;

[0054] A first calculation unit for calculating the structural parameter fitting coefficients corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber;

[0055] A second construction unit for constructing an intelligent inverse design optimization model of the bend-resistant photonic crystal fiber not less than the structural parameter fitting coefficients and the minimum fiber bend-resistant performance parameter among the corresponding second-dimensional fiber bend-resistant performance parameters.

[0056] According to one aspect of the present disclosure, there is provided an intelligent design system for bend-resistant photonic crystal fibers, applied to reverse prediction tasks, including: an electronic device configured with a processor and a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the above intelligent design method.

[0057] According to one aspect of the present disclosure, there is provided an intelligent design system for bend-resistant photonic crystal fibers, applied to reverse prediction tasks, including: a computer-readable storage medium storing computer program instructions thereon, and the computer program instructions implement the above intelligent design method when executed by a processor.

[0058] According to one aspect of the present disclosure, there is provided an intelligent design system for bend-resistant photonic crystal fibers, applied to reverse prediction tasks, including: a computer program product provided with a computer program / instructions, and the computer program / instructions implement the intelligent design method as above when executed by a processor.

[0059] According to one aspect of the present disclosure, there is provided an intelligent design of bend-resistant photonic crystal fibers, applied to forward prediction tasks, including: the intelligent design system of bend-resistant photonic crystal fibers applied to reverse prediction tasks as above; and,

[0060] A third construction unit for constructing multiple intelligent forward design models of bend-resistant photonic crystal fibers for predicting the second-dimensional fiber bend-resistant performance parameters based on the first-dimensional structural parameters and the corresponding second-dimensional fiber bend-resistant performance parameters, using the preset machine learning algorithm or preset network corresponding to the intelligent design of bend-resistant photonic crystal fibers;

[0061] A second calculation unit for respectively calculating multiple fiber bend-resistant performance fitting coefficients corresponding to the multiple intelligent forward design models of bend-resistant photonic crystal fibers;

[0062] A determination unit for determining an intelligent optimal forward design model of bend-resistant photonic crystal fibers based on the multiple fiber bend-resistant performance fitting coefficients.

[0063] According to one aspect of the present disclosure, there is provided an intelligent design of a bend-resistant photonic crystal fiber, which is applied to a forward prediction task, including: an electronic device configured with a processor and a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the above-mentioned intelligent design method.

[0064] According to one aspect of the present disclosure, there is provided an intelligent design of a bend-resistant photonic crystal fiber, which is applied to a forward prediction task, including: a computer-readable storage medium storing computer program instructions thereon, and when the computer program instructions are executed by a processor, the above-mentioned intelligent design method is implemented.

[0065] According to one aspect of the present disclosure, there is provided a computer program product of an intelligent design of a bend-resistant photonic crystal fiber, which is applied to a forward prediction task, including: the computer program product is provided with a computer program / instructions, and when the computer program / instructions are executed by a processor, the above-mentioned intelligent design method is implemented.

[0066] According to one aspect of the present disclosure, there is provided an intelligent analysis system of a bend-resistant photonic crystal fiber, including: the intelligent design system of a bend-resistant photonic crystal fiber applied to a reverse prediction task as described above; and,

[0067] A locality interpretability unit for using a locality interpretability algorithm to configure each fiber bend-resistant performance parameter corresponding to the minimum-dimension fiber bend-resistant performance parameter in the test set for training the intelligent reverse design optimization model of the bend-resistant photonic crystal fiber as a specific sample;

[0068] A perturbation sample generation unit for using an interpretation instance function to adjust the eigenvalue corresponding to the specific sample to generate a corresponding perturbation sample;

[0069] A contribution degree determination unit for determining the contribution degree to the first-dimensional structural parameter determined by the locality interpretability algorithm based on the intelligent reverse design optimization model of the bend-resistant photonic crystal fiber, the specific sample, the perturbation sample, and the locality interpretability algorithm;

[0070] A judgment unit for determining that the minimum-dimension fiber bend-resistant performance parameter corresponding to the intelligent reverse design optimization model of the bend-resistant photonic crystal fiber is optimal if the contribution degree to the first-dimensional structural parameter determined by the locality interpretability algorithm is basically consistent with the contribution degree to the first-dimensional structural parameter determined by the global interpretability algorithm.

[0071] According to one aspect of the present disclosure, there is provided an intelligent analysis system for a bend-resistant photonic crystal fiber, including: an electronic device configured with a processor and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the above-mentioned intelligent analysis method.

[0072] According to one aspect of the present disclosure, there is provided a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the intelligent analysis method as described above is implemented.

[0073] According to one aspect of the present disclosure, there is provided a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the intelligent analysis method as described above is implemented.

[0074] In an embodiment of the present disclosure, a technical solution corresponding to an intelligent design, analysis method, and system for a bend-resistant photonic crystal fiber proposed by the present disclosure is provided to solve the problems in the traditional design method of bend-resistant PCF that the calculation resources of structural parameters consume a large amount, resulting in a long design cycle of bend-resistant PCF, and serious deficiencies in the interpretability of the structural parameters of bend-resistant PCF and the transparency of the design process of the intelligent design model of bend-resistant PCF.

[0075] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and do not limit the present disclosure.

[0076] According to the following detailed description of exemplary embodiments with reference to the accompanying drawings, other features and aspects of the present disclosure will become clear. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and, together with the specification, are used to explain the technical solutions of the present disclosure.

[0078] Figure 1 A flowchart showing an intelligent design method for a bend-resistant photonic crystal fiber applied to a reverse prediction task according to an embodiment of the present disclosure;

[0079] Figure 2 A cross-sectional view of a bend-resistant PCF showing an embodiment of the present disclosure;

[0080] Figure 3 A network structure diagram corresponding to a DNN network showing an embodiment of the present disclosure;

[0081] Figure 4Shows the loss function curves on the training set and test set based on the DNN network according to an embodiment of the present disclosure;

[0082] Figure 5 Shows the prediction effect on the test set of the intelligent inverse design model of the bending-resistant photonic crystal fiber based on the DNN network according to an embodiment of the present disclosure; wherein, (a) air hole spacing; (b) the second diameter (small air hole diameter) of the air holes corresponding to the second size smaller than the first size; (c) the first diameter (large air hole diameter) of the air holes corresponding to the first size;

[0083] Figure 6 Shows the sorted summary diagram of the feature importance corresponding to three structural parameters output by the intelligent design model of the bending-resistant photonic crystal fiber based on the DNN network under the SHAP algorithm according to an embodiment of the present disclosure;

[0084] Figure 7 Shows the bending performance at different wavelengths and air hole spacings according to an embodiment of the present disclosure; wherein, (a) the first bending loss and the second bending loss; (b) the effective mode area;

[0085] Figure 8 Shows the bending performance at different wavelengths and large air hole diameters d2 according to an embodiment of the present disclosure; wherein, (a) the bending loss; (b) the effective mode area;

[0086] Figure 9 Shows the bending performance at different wavelengths and small air hole diameters (second diameter d1) according to an embodiment of the present disclosure; wherein, (a) the first bending loss and the second bending loss; (b) the effective mode area;

[0087] Figure 10 Shows the loss function curves on the training set and validation set based on the DNN network and the bending-resistant performance parameters of the second-dimensional fiber according to an embodiment of the present disclosure;

[0088] Figure 11 Shows the loss function curves on the training set and validation set based on the optimized DNN network and the bending-resistant performance parameters of the minimum-dimensional fiber according to an embodiment of the present disclosure;

[0089] Figure 12 Shows the prediction effect and training time based on the machine learning algorithm according to an embodiment of the present disclosure;

[0090] Figure 13 Shows the influence analysis diagram of three structural parameters under the LIME algorithm according to an embodiment of the present disclosure. Detailed implementation manners

[0091] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. Like reference numerals in the drawings denote functionally identical or similar elements. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0092] As used herein, the term "exemplary" means "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as superior to or better than other embodiments.

[0093] As used herein, the term "and / or" merely describes an associative relationship between associated objects and represents three relationships that may exist. For example, A and / or B may represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the term "at least one" as used herein represents any one of a plurality or any combination of at least two of a plurality. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set consisting of A, B, and C.

[0094] In addition, for a better illustration of the present disclosure, numerous specific details are given in the following detailed description. Those skilled in the art should understand that the present disclosure can be implemented without some specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present disclosure.

[0095] It can be understood that the above-described embodiments of the intelligent design and analysis methods of the bending-resistant photonic crystal fiber according to the present disclosure can be combined with each other to form a combined embodiment without violating the principle logic. Due to space limitations, the present disclosure will not elaborate further.

[0096] In addition, the present disclosure also provides an intelligent design and analysis device or system, an electronic device, a computer-readable storage medium, and a program product for the bending-resistant photonic crystal fiber. The above can all be used to implement any one of the intelligent design and analysis methods of the bending-resistant photonic crystal fiber provided by the present disclosure. The corresponding technical solutions and descriptions are referred to the corresponding records in the method section and will not be elaborated further.

[0097] Figure 1 The flowchart showing the intelligent design method of the bending-resistant photonic crystal fiber according to an embodiment of the present disclosure is as Figure 1As shown, the intelligent design method of the bend-resistant photonic crystal fiber includes: Step S101: Generate the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the second-dimensional fiber bend-resistant performance parameters corresponding to the dimensions greater than the first-dimensional structural parameters; Step S102: Based on the first-dimensional structural parameters and the corresponding second-dimensional fiber bend-resistant performance parameters, construct an intelligent inverse design model of the bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters; Step S103: Calculate the structural parameter fitting coefficient corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber; Step S104: Construct an intelligent inverse design optimization model of the bend-resistant photonic crystal fiber not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter among the corresponding second-dimensional fiber bend-resistant performance parameters. To solve the problems in the traditional design method of bend-resistant PCF that the structural parameter calculation consumes a large amount of resources, resulting in a long design cycle of bend-resistant PCF, and there are serious deficiencies in the interpretability of the structural parameters of bend-resistant PCF and the transparency of the design process of the intelligent design model of bend-resistant PCF.

[0098] Step S101: Generate the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the second-dimensional fiber bend-resistant performance parameters corresponding to the dimensions greater than the first-dimensional structural parameters.

[0099] In an embodiment of the present disclosure, the method for generating the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the second-dimensional fiber bend-resistant performance parameters corresponding to the dimensions greater than the first-dimensional structural parameters includes: Using physical field simulation software, under the set working wavelength and the first-dimensional structural parameters corresponding to the bend-resistant photonic crystal fiber, adjust the bending radius corresponding to the photonic crystal fiber to generate the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the second-dimensional fiber bend-resistant performance parameters corresponding to the dimensions greater than the first-dimensional structural parameters.

[0100] In an embodiment of the present disclosure, the method for adjusting the bending radius of the photonic crystal fiber at a set operating wavelength and the first-dimensional structural parameters corresponding to the bend-resistant photonic crystal fiber to generate the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the corresponding second-dimensional fiber bend-resistant performance parameters thereof includes: Step S201: Configure the first-dimensional structural parameters as the first diameter of the air holes corresponding to the first size in the bend-resistant photonic crystal fiber, the second diameter of the air holes corresponding to the second size smaller than the first size, and the air hole pitch between the air holes; Step S202: Obtain the set ranges and set step sizes corresponding to the first diameter, the second diameter, the air hole pitch, and the bending radius; Step S203: Configure any one of the structural parameters among the first diameter, the second diameter, and the air hole pitch as a variable parameter, and configure the structural parameters other than the variable parameter as constant parameters; Step S204: Starting from any endpoint of the set range corresponding to the variable parameter, at the set operating wavelength, adjust the bending radius of the photonic crystal fiber, and determine the parameter change corresponding to the variable parameter according to the corresponding set step; Step S205: Determine the corresponding second-dimensional fiber bend-resistant performance parameters under the parameter change corresponding to the variable parameter and the constant parameters; Loop or repeat the above steps S203-S205 until the first diameter, the second diameter, and the air hole pitch are configured as variable parameters, thereby generating the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the corresponding second-dimensional fiber bend-resistant performance parameters thereof.

[0101] In the embodiments of the present disclosure and other possible embodiments, the second-dimensional fiber bend-resistant performance parameters include one or several of the first bending loss of the high-order mode, the second bending loss of the fundamental mode, and the effective mode area corresponding to the bend-resistant photonic crystal fiber under the first-dimensional structural parameters or different structural parameters.

[0102] Figure 2 Shows a cross-sectional view of a bend-resistant PCF according to an embodiment of the present disclosure. As Figure 2 shown, when the bending radius of the bend-resistant PCF is 10 cm, the single-mode bandwidth of the bend-resistant PCF is 1.1 μm - 2.8 μm and the effective mode area of the bend-resistant PCF can reach 1055.3 μm 2 The cross-section of the broadband width, large mode field, and corresponding single-mode bend-resistant PCF (referred to as bend-resistant PCF for short). Among them, the parameters d1, d2, and Λ respectively represent the second diameter of the air holes corresponding to the second size in the PCF, the first diameter of the air holes corresponding to the first size larger than the second size, and the air hole pitch between the air holes.

[0103] As Figure 2As shown, in the embodiments of the present disclosure and other possible embodiments, an optical fiber core is disposed at the center of the bend-resistant PCF. After removing the air holes corresponding to the four dotted lines near the optical fiber core, the remaining air holes form a three-ring hexagonal structure. In the innermost ring (the first ring) of the bend-resistant PCF, multiple ( Figure 2 six in number) air holes corresponding to the first diameter d2 of the largest first size confine light within the optical fiber core, thereby suppressing the distortion of the bend-resistant PCF caused by bending.

[0104] In the embodiments of the present disclosure and other possible embodiments, the second ring outside the first ring corresponding to the direction from the optical fiber core to the Perfectly Matched Layers (PML) contains multiple ( Figure 2 fourteen in number) air holes, and the third ring outside the second ring corresponding to the direction from the optical fiber core to the PML contains multiple ( Figure 2 sixteen in number) air holes. The outermost ring outside the third ring corresponding to the direction from the optical fiber core to the PML removes multiple ( Figure 2 four in number) air holes that are symmetric at the upper and lower ends, increasing the leakage of higher-order modes to a certain extent and maintaining the single-mode characteristics of the bend-resistant PCF.

[0105] In addition, in the embodiments of the present disclosure and other possible embodiments, the bend-resistant PCF only includes circular air holes corresponding to two sizes (the first size and the second size) and a background material; the background material can be configured as arsenic trisulfide As2S3, and arsenic trisulfide As2S3 is distributed between the circular air holes corresponding to the two sizes (the first size and the second size) and inside the PML, facilitating the preparation of the bend-resistant PCF.

[0106] In addition, in the embodiments of the present disclosure and other possible embodiments, when the background material of this bend-resistant PCF is configured as chalcogenide glass As2S3, the refractive index n(λ) of the background material varies with the wavelength λ of the incident light and can be calculated by the Sellmeier equation:

[0107]

[0108] where A n and B n are the Sellmeier coefficients, λ is the wavelength of the incident light, and the unit is μm. Among them, the refractive index n(λ) of the background material directly affects the performance parameters of the bend-resistant PCF. The Sellmeier coefficients of As2S3 are shown in Table 1.

[0109]

[0110] Table 1 Sellmeier coefficients of As2S3

[0111] In the embodiments of the present disclosure and other possible embodiments, the core task of generating a dataset of anti-bending PCF corresponding to the first-dimensional structural parameters of the anti-bending photonic crystal fiber and the anti-bending performance parameters of the second-dimensional fiber corresponding to the dimensions greater than the first-dimensional structural parameters is to construct a comprehensive and accurate dataset of anti-bending PCF. Based on the structural parameters such as air hole spacing and air hole diameter (fiber structural parameters) corresponding to different structural parameters and their set ranges and set change steps, and under different influencing factors such as operating wavelength and bending radius and their set ranges and set change steps, a physical field simulation software is used to generate data corresponding to the structural parameters of the anti-bending PCF and the fiber anti-bending performance parameters corresponding to the structural parameters one by one.

[0112] In the embodiments of the present disclosure and other possible embodiments, in the reverse prediction task, using a physical field simulation software (COMSOL Multiphysics software), at a set operating wavelength of 1.31 μm and different structural parameters corresponding to the anti-bending PCF, the bending radius is adjusted to obtain the fiber anti-bending performance parameters of 771 * 33 dimensions (second dimension) corresponding to the 771 * 3 dimensions (first dimension) of structural parameters. Among them, the dimension of the first dimension is configured as 3, and the dimension of the second dimension is configured as 33.

[0113] In the embodiments of the present disclosure and other possible embodiments, the set ranges and set steps corresponding to the first diameter, the second diameter, and the air hole spacing are respectively configured as a third set range, a second set range, a first set range, a third set step, a second set step, and a first set step. Among them, the set range and set step corresponding to the bending radius are respectively configured within a fourth set range (for example, 10 - 20 cm) and a fourth set step (for example, 1 cm).

[0114] Specifically, in the embodiments of the present disclosure and other possible embodiments, different structural parameters (first-dimensional structural parameters) corresponding to the anti-bending PCF include: three structural parameters d1, d2, and Λ; among them, the parameters d1, d2, and Λ respectively represent the first diameter of the air holes corresponding to the first size in the anti-bending PCF, the second diameter of the air holes corresponding to the second size smaller than the first size, and the air hole spacing between the air holes. Structure The third set range, the second set range, and the first set range configured by the parameters d1, d2, and Λ are respectively configured as 9 - 11 μm, 19 - 23 μm, 23 - 25.4 μm, and the third set step, the second set step, and the first set step (change step) corresponding to the third set range, the second set range, and the first set range are respectively configured as 0.2 μm, to obtain the corresponding 771 * 3 dimensions of structural parameters (771 * first-dimensional structural parameters).

[0115] Specifically, in the embodiments of the present disclosure and other possible embodiments, within the fourth set range (e.g., 10 - 20 cm) of the bending radius of the bend-resistant PCF and corresponding to the fourth set step size (e.g., 1 cm), the second-dimensional fiber bend-resistant performance parameters of the second-dimensional optical fiber include: the first bending loss of 11 HOMs, the second bending loss of 11 FMs, and the effective mode area of 11. At the set working wavelength of 1.31 μm, under different structural parameters corresponding to the bend-resistant PCF, and at 11 different bending radii, 771 * 3-dimensional performance parameters corresponding to 771 * 33 (771 * 11 * 3)-dimensional fiber bend-resistant performance parameters (771 * second-dimensional fiber bend-resistant performance parameters) are obtained.

[0116] Step S102: Based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters, construct an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters.

[0117] In the embodiments of the present disclosure, before constructing an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters, logarithmic processing is performed on the first bending loss of the high-order modes and the second bending loss of the fundamental mode in the second-dimensional fiber bend-resistant performance parameters to obtain the logarithmic first bending loss and the logarithmic second bending loss; normalization processing is performed on the logarithmic first bending loss, the logarithmic second bending loss, and the effective mode area in the second-dimensional fiber bend-resistant performance parameters.

[0118] In addition, in the embodiments of the present disclosure and other possible embodiments, logarithmic processing with base 10 is performed on both the first bending loss and the second bending loss to obtain the logarithmic first bending loss and the logarithmic second bending loss; normalization preprocessing is performed on the structural parameters and the fiber bend-resistant performance parameters (including: the logarithmic first bending loss and the logarithmic second bending loss) to obtain the normalized structural parameters and the normalized fiber bend-resistant performance parameters.

[0119] In the embodiments of the present disclosure, the method for constructing an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters includes: determining a training set based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters; using the training set to train a preset machine learning algorithm or a preset network corresponding to the intelligent design of a bend-resistant photonic crystal fiber, and constructing an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters.

[0120] In the embodiments of the present disclosure and other possible embodiments, the preset machine learning algorithm or preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber can be configured as one or several of support vector regression (SVR), support vector machine (SVM), K-nearest neighbor algorithm (KNN), ensemble algorithm, XGBoost, random forest (RF), artificial neural network (ANN), deep neural network (DNN), etc.

[0121] In summary, in the embodiments of the present disclosure and other possible embodiments, a total of 771×33-dimensional input corresponding to the second-dimensional optical fiber bend-resistant performance parameters are constructed; for each of the 771 dimensions, 1 group of structural parameters corresponding to 33 second-dimensional optical fiber bend-resistant performance parameters; among them, each group of structural parameters includes three structural parameters d1, d2, and Λ. Among them, each group of data in the bend-resistant PCF dataset corresponding to the bend-resistant PCF includes 33 second-dimensional optical fiber bend-resistant performance parameters and their corresponding 3 structural parameters. The bend-resistant PCF dataset corresponding to the bend-resistant PCF is randomly divided into a training set and a test set according to a set ratio of 3:1.

[0122] In the embodiments of the present disclosure and other possible embodiments, in order to avoid the non-uniqueness problem in the inverse prediction task, when constructing the bend-resistant PCF dataset corresponding to the inverse prediction task, the working wavelength of the bend-resistant PCF is configured as a set working wavelength (for example, configured as 1.31 μm). In the bend-resistant PCF dataset corresponding to the inverse prediction task, the input for the DNN includes the first bending loss of the HOM, the second bending loss of the FM, and the effective mode area and their corresponding structural parameters. Among them, the first bending loss of the HOM, the second bending loss of the FM, and the effective mode area respectively include the second-dimensional optical fiber bend-resistant performance parameters (the first bending loss of the HOM, the second bending loss of the FM, and the effective mode area) corresponding to the fourth set range (for example, 10 - 20 cm) of the bending radius of the bend-resistant PCF and the fourth set step length (for example, 1 cm). Among them, the three numerical values (11 first bending losses of the HOM, 11 second bending losses of the FM, and 11 effective mode areas) are arranged in sequence to jointly form the second-dimensional optical fiber bend-resistant performance parameters corresponding to the 11×3-dimensional (3 groups) input.

[0123] Step S103: Calculate the structural parameter fitting coefficient corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber.

[0124] In an embodiment of the present disclosure, the method for calculating the structural parameter fitting coefficient corresponding to the anti-bending photonic crystal fiber intelligent inverse design model includes: determining a test set based on the first-dimensional structural parameters and the corresponding second-dimensional fiber anti-bending performance parameters; configuring the performance index corresponding to the anti-bending photonic crystal fiber intelligent inverse design model as the structural parameter fitting coefficient; and using the test set to calculate the structural parameter fitting coefficient corresponding to the anti-bending photonic crystal fiber intelligent inverse design model.

[0125] In the embodiments of the present disclosure and other possible embodiments, the output of the anti-bending PCF intelligent design model based on the DNN network includes: the structural parameters corresponding to the second-dimensional fiber anti-bending performance parameters. For example, the first-dimensional structural parameters corresponding to the second-dimensional fiber anti-bending performance parameters include: three structural parameters d1, d2, and Λ; where the parameters d1, d2, and Λ respectively represent the diameter of the air holes corresponding to the first dimension in the anti-bending PCF, the diameter of the air holes corresponding to the second dimension smaller than the first dimension, and the air hole pitch between the air holes.

[0126] In the embodiments of the present disclosure and other possible embodiments, the data set corresponding to the anti-bending PCF can be described as the change spectrum of the fiber anti-bending performance parameters of the anti-bending PCF for a specific structure with respect to the bending radius at a set working wavelength of 1.31 μm. The DNN network is to find the mapping relationship between the fiber anti-bending performance parameters of the anti-bending PCF and its corresponding structural parameters. The input of the network is the fiber anti-bending performance parameters of the PCF corresponding to the optical characteristics in the bent state, and the output is the structural parameters corresponding to the optical structure mapped to the fiber anti-bending performance parameters of the PCF. Through this DNN network, the mapping relationship between the fiber anti-bending performance parameters of the above anti-bending PCF and its corresponding structural parameters is found.

[0127] Step S104: Construct an anti-bending photonic crystal fiber intelligent inverse design optimization model not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber anti-bending performance parameter in the corresponding second-dimensional fiber anti-bending performance parameters.

[0128] In the embodiments of the present disclosure and other possible embodiments, the XAI technology is introduced in the technical solution, which can analyze the interpretability of the anti-bending PCF intelligent design model based on the DNN, and then perform feature selection corresponding to the second-dimensional fiber anti-bending performance parameters and optimization of the anti-bending PCF intelligent design model (anti-bending photonic crystal fiber intelligent inverse design model).

[0129] In an embodiment of the present disclosure, the method for constructing an anti-bending photonic crystal fiber intelligent reverse design optimization model not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber anti-bending performance parameter among the corresponding second-dimensional fiber anti-bending performance parameters includes: using a global interpretability algorithm to interpret the second-dimensional fiber anti-bending performance parameters, each fiber anti-bending performance parameter of each dimension of the second-dimensional fiber anti-bending performance parameters; sorting the second-dimensional fiber anti-bending performance parameters based on the contribution degree corresponding to each fiber anti-bending performance parameter to obtain the sorted second-dimensional fiber anti-bending performance parameters; selecting the corresponding updated-dimensional fiber anti-bending performance parameters from the sorted second-dimensional fiber anti-bending performance parameters according to the contribution degree corresponding to each fiber anti-bending performance parameter and a set contribution degree; and constructing an anti-bending photonic crystal fiber intelligent reverse design optimization model not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber anti-bending performance parameter among the corresponding second-dimensional fiber anti-bending performance parameters based on the updated-dimensional fiber anti-bending performance parameters, the first-dimensional structural parameters, and the corresponding optimized preset machine learning algorithm or preset network.

[0130] In an embodiment of the present disclosure, the method for constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the minimum bend-resistant performance parameter of the second-dimensional fiber bend-resistant performance parameter corresponding to the structure parameter fitting coefficient, based on the updated-dimensional fiber bend-resistant performance parameter, the first-dimensional structure parameter, and the corresponding optimized preset machine learning algorithm or preset network, includes: Step 301: Optimize the preset machine learning algorithm or preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber using the updated-dimensional fiber bend-resistant performance parameter corresponding to the second-dimensional fiber bend-resistant performance parameter; Step 302: Train the optimized preset machine learning algorithm or preset network using the updated-dimensional fiber bend-resistant performance parameter and its corresponding first-dimensional structure parameter; Step 303: Calculate the optimized structure parameter fitting coefficient of the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber corresponding to the optimized preset machine learning algorithm or preset network; Step 304: Compare the optimized structure parameter fitting coefficient with the structure parameter fitting coefficient corresponding to the second-dimensional fiber bend-resistant performance parameter; Step 304: If the optimized structure parameter fitting coefficient is greater than or equal to the structure parameter fitting coefficient corresponding to the second-dimensional fiber bend-resistant performance parameter, obtain the set deletion dimension; Delete the fiber bend-resistant performance parameter corresponding to the set deletion dimension with a lower ranking from the updated-dimensional fiber bend-resistant performance parameter to obtain the latest updated-dimensional fiber bend-resistant performance parameter; where the set deletion dimension is greater than or equal to 1 and is a positive integer; Based on the latest updated-dimensional fiber bend-resistant performance parameter, repeat Steps 301 to 304 until an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the minimum bend-resistant performance parameter of the second-dimensional fiber bend-resistant performance parameter corresponding to the structure parameter fitting coefficient is constructed; Step 305: If the optimized structure parameter fitting coefficient is less than the structure parameter fitting coefficient corresponding to the second-dimensional fiber bend-resistant performance parameter, obtain the set addition dimension; Add the fiber bend-resistant performance parameter corresponding to the set deletion dimension with a higher ranking from the updated-dimensional fiber bend-resistant performance parameter to obtain the latest updated-dimensional fiber bend-resistant performance parameter; where the set addition dimension is greater than or equal to 1 and is a positive integer; Based on the latest updated-dimensional fiber bend-resistant performance parameter, repeat Steps 301 to 303 and Step 305 until an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the minimum bend-resistant performance parameter of the second-dimensional fiber bend-resistant performance parameter corresponding to the structure parameter fitting coefficient is constructed.

[0131] In the embodiments of the present disclosure and other possible embodiments, the method for constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the minimum dimension fiber bend-resistant performance parameter among the second dimension fiber bend-resistant performance parameters corresponding to the structural parameter fitting coefficient based on the latest updated dimension fiber bend-resistant performance parameter includes: Step 301: Optimize the preset machine learning algorithm or preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber by using the latest updated dimension fiber bend-resistant performance parameter corresponding to the second dimension fiber bend-resistant performance parameter; Step 302: Train the optimized preset machine learning algorithm or preset network by using the latest updated dimension fiber bend-resistant performance parameter and its corresponding first dimension structural parameter; Step 303: Calculate the optimized structural parameter fitting coefficient of the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber corresponding to the optimized preset machine learning algorithm or preset network; Step 304: Compare the optimized structural parameter fitting coefficient with the structural parameter fitting coefficient corresponding to the second dimension fiber bend-resistant performance parameter; Step 304: If the optimized structural parameter fitting coefficient is greater than or equal to the structural parameter fitting coefficient corresponding to the second dimension fiber bend-resistant performance parameter, obtain a set of deleted dimensions; Delete the fiber bend-resistant performance parameter corresponding to the set of deleted dimensions with a lower ranking from the latest updated dimension fiber bend-resistant performance parameter to obtain the latest updated dimension fiber bend-resistant performance parameter.

[0132] In the embodiments of the present disclosure and other possible embodiments, the method for constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber that is not lower than the minimum dimension fiber bend-resistant performance parameter among the second dimension fiber bend-resistant performance parameters corresponding to the structural parameter fitting coefficient based on the latest updated dimension fiber bend-resistant performance parameter, includes: Step 301: Optimize the preset machine learning algorithm or preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber by using the latest updated dimension fiber bend-resistant performance parameter corresponding to the second dimension fiber bend-resistant performance parameter; Step 302: Train the optimized preset machine learning algorithm or preset network by using the latest updated dimension fiber bend-resistant performance parameter and its corresponding first dimension structural parameter; Step 303: Calculate the optimized structural parameter fitting coefficient of the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber corresponding to the optimized preset machine learning algorithm or preset network; Step 305: If the optimized structural parameter fitting coefficient is less than the structural parameter fitting coefficient corresponding to the second dimension fiber bend-resistant performance parameter, obtain a set of added dimensions; Add the fiber bend-resistant performance parameter corresponding to the set of added dimensions with a higher ranking from the latest updated dimension fiber bend-resistant performance parameter to obtain the latest updated dimension fiber bend-resistant performance parameter.

[0133] In embodiments of the present disclosure and other possible embodiments, those skilled in the art can configure the set deletion dimension and the set addition dimension respectively according to actual needs. For example, the set deletion dimension and the set addition dimension can be configured as 1 or other values.

[0134] In embodiments of the present disclosure and other possible embodiments, based on the contribution degree corresponding to each optical fiber bending resistance performance parameter, the second-dimensional optical fiber bending resistance performance parameters are sorted from large to small, and the sorted second-dimensional optical fiber bending resistance performance parameters are configured as X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, …, X35, …, X40; according to the contribution degree corresponding to each optical fiber bending resistance performance parameter and the set contribution degree, the updated-dimensional optical fiber bending resistance performance parameters greater than the set contribution degree are selected from the sorted second-dimensional optical fiber bending resistance performance parameters, and the updated-dimensional optical fiber bending resistance performance parameters X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, …, X35 are obtained.

[0135] In embodiments of the present disclosure and other possible embodiments, when the set deletion dimension is configured as 1, step 304 is executed for the first time: if the optimization structure parameter fitting coefficient is greater than or equal to the structure parameter fitting coefficient corresponding to the second-dimensional optical fiber bending resistance performance parameter, then the set deletion dimension is obtained; the optical fiber bending resistance performance parameter corresponding to the set deletion dimension with a later ranking is deleted from the updated-dimensional optical fiber bending resistance performance parameters, and the latest updated-dimensional optical fiber bending resistance performance parameters X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, …, X34 are obtained.

[0136] In embodiments of the present disclosure and other possible embodiments, the set deletion dimension and the set addition dimension can be configured as 1, and step 305 is executed for the first time: if the optimization structure parameter fitting coefficient is less than the structure parameter fitting coefficient corresponding to the second-dimensional optical fiber bending resistance performance parameter, then the set addition dimension is obtained; the optical fiber bending resistance performance parameter corresponding to the set deletion dimension with a higher ranking is added to the updated-dimensional optical fiber bending resistance performance parameters, and the latest updated-dimensional optical fiber bending resistance performance parameters X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, …, X35, X36 are obtained.

[0137] In an embodiment of the present disclosure, the method for optimizing a corresponding preset machine learning algorithm or preset network for intelligent design of the bend-resistant photonic crystal fiber by using the updated dimension fiber bend-resistant performance parameters includes: when selecting the corresponding updated dimension fiber bend-resistant performance parameters from the sorted second dimension fiber bend-resistant performance parameters or deleting the fiber bend-resistant performance parameters corresponding to the set deletion dimensions with lower rankings from the updated dimension fiber bend-resistant performance parameters to obtain the latest updated dimension fiber bend-resistant performance parameters, adjusting the number of neurons corresponding to the input layer of the preset machine learning algorithm or preset network for intelligent design of the bend-resistant photonic crystal fiber by using the updated dimension fiber bend-resistant performance parameters or the latest updated dimension fiber bend-resistant performance parameters; reducing the number of layers of the middle hidden layer and the number of neurons in each layer of the preset machine learning algorithm or preset network for intelligent design of the bend-resistant photonic crystal fiber according to the set layer step size to obtain the optimized preset machine learning algorithm or preset network; wherein, the number of neurons corresponding to the input layer is the same as the dimension corresponding to the updated dimension fiber bend-resistant performance parameters or the latest updated dimension fiber bend-resistant performance parameters; or, when adding the fiber bend-resistant performance parameters corresponding to the set deletion dimensions with higher rankings from the updated dimension fiber bend-resistant performance parameters to obtain the latest updated dimension fiber bend-resistant performance parameters, adjusting the number of neurons corresponding to the input layer of the preset machine learning algorithm or preset network for intelligent design of the bend-resistant photonic crystal fiber by using the updated dimension fiber bend-resistant performance parameters; increasing the number of layers of the middle hidden layer and the number of neurons in each layer of the preset machine learning algorithm or preset network for intelligent design of the bend-resistant photonic crystal fiber according to the set layer step size to obtain the optimized preset machine learning algorithm or preset network; wherein, the number of neurons corresponding to the input layer is the same as the dimension corresponding to the latest updated dimension fiber bend-resistant performance parameters.

[0138] In the embodiments of the present disclosure and other possible embodiments, those skilled in the art can configure the set layer step size according to actual needs. For example, the set layer step size is configured as 1 or other values.

[0139] In embodiments of the present disclosure and other possible embodiments, the optimized intermediate hidden layer of the preset machine learning algorithm or preset network includes: a first intermediate hidden layer connected to the input layer, a plurality of second intermediate hidden layers, and a last intermediate hidden layer connected to the output layer of the preset machine learning algorithm or preset network corresponding to the intelligent design of the anti-bending photonic crystal fiber; wherein, a first activation function is configured after each neuron of the first intermediate hidden layer; a regularization function is configured after each neuron of each second intermediate hidden layer in the plurality of second intermediate hidden layers; a second activation function is configured after each neuron of the last intermediate hidden layer; wherein, the input of the first intermediate hidden layer among the plurality of second intermediate hidden layers is connected to the first intermediate hidden layer, and the output of the first second intermediate hidden layer is connected to the input of the second second intermediate hidden layer; and so on, the output layer of the last second intermediate hidden layer is connected to the input of the last intermediate hidden layer; wherein, the first activation function is configured as a ReLU activation function; and the second activation function is configured as a tanh activation function.

[0140] In the embodiments of the present disclosure and other possible embodiments, the process of the inverse design scheme (reverse prediction task, i.e., the input is the fiber bending resistance performance parameter and the output is the structural parameter) is as follows: First, using a physical field simulation software (COMSOL Multiphysics software), at the set working wavelength and different structural parameters corresponding to the bend-resistant PCF, adjust the bending radius to obtain the fiber bending resistance performance parameters corresponding to the multi-dimensional structural parameters, and generate a bend-resistant PCF dataset corresponding to the reverse prediction task; then use the generated bend-resistant PCF dataset to train the DNN network, and preliminarily evaluate the bend-resistant PCF intelligent design model based on the DNN and its prediction performance; then, use XAI technology to deeply analyze the key fiber bending resistance performance parameters of the fiber bending resistance performance parameters in the bend-resistant PCF intelligent design model based on the DNN, identify and eliminate the fiber bending resistance performance parameters with less contribution to the predicted structural parameters; finally, analyze the importance corresponding to the global fiber bending resistance performance interpretation parameters and / or local fiber bending resistance performance interpretation parameters based on XAI technology, and determine the optimized third-dimensional fiber bending resistance performance parameter (updated dimension fiber bending resistance performance parameter or latest updated dimension fiber bending resistance performance parameter or minimum dimension fiber bending resistance performance parameter) corresponding to the key fiber bending resistance performance parameter; based on the optimized third-dimensional fiber bending resistance performance parameter (updated dimension fiber bending resistance performance parameter or latest updated dimension fiber bending resistance performance parameter or minimum dimension fiber bending resistance performance parameter), adjust the input layer of the bend-resistant PCF intelligent design network so that the number of the input layer of the bend-resistant PCF intelligent design network matches the number corresponding to the optimized third-dimensional fiber bending resistance performance parameter (updated dimension fiber bending resistance performance parameter or latest updated dimension fiber bending resistance performance parameter or minimum dimension fiber bending resistance performance parameter), and obtain the optimized bend-resistant PCF intelligent design network; use the optimized third-dimensional fiber bending resistance performance parameter (updated dimension fiber bending resistance performance parameter or latest updated dimension fiber bending resistance performance parameter or minimum dimension fiber bending resistance performance parameter) to retrain the optimized bend-resistant PCF intelligent design network to further improve its prediction performance and generalization ability.

[0141] Figure 3 Shows the network structure diagram corresponding to the DNN network according to the embodiment of the present disclosure. The DNN network used in the present invention is a fully connected neural network, and its structure is as Figure 3As shown. The input layer and output layer of the DNN model network respectively contain a first number (e.g., 33) of neurons corresponding to the number of second-dimensional fiber bending resistance performance parameters and a second number (e.g., 3) of neurons corresponding to the number of structural parameters. There is an intermediate hidden layer, which contains 4 layers; each layer respectively contains a third number (e.g., 50) of neurons. To overcome the vanishing gradient problem and accelerate learning, in addition to the Sigmoid and Tanh activation functions, the ReLU activation function is also used to approximate the non-linear function. Specifically, the ReLU activation function is configured after each neuron in the second layer of the intermediate hidden layer and after each neuron in the fourth layer, and the Sigmoid activation function and Tanh activation function are respectively after each neuron in the first layer and the third layer of the intermediate hidden layer.

[0142] This is because during the training process of deep neural networks, the vanishing gradient is a common problem. For the Sigmoid and Tanh activation functions, when the input value is large or small, the gradient will approach 0, resulting in slow or even stagnant update of the DNN network parameters, affecting the training effect of the bending-resistant photonic crystal fiber intelligent design model. The ReLU activation function (Rectified Linear Unit) has a constant gradient of 1 in the positive part, which can effectively avoid the vanishing gradient problem, accelerate the convergence of the DNN network, and make the bending-resistant photonic crystal fiber intelligent design model learn faster. After each neuron in each hidden layer, a batch normalization layer is added to reduce the internal covariate shift. In the neurons of the penultimate hidden layer (the third hidden layer), Dropout regularization with a set value (0.1) is applied. In addition, Adam is used as the optimizer, and the Adam optimizer performs well when dealing with relatively large datasets. And the mean square error (MSE) is used as the loss function calculated based on the difference between the actual value (the multi-dimensional structural parameters for generating the bending-resistant photonic crystal fiber) and the predicted value (the multi-dimensional structural parameters of the bending-resistant photonic crystal fiber predicted by the bending-resistant photonic crystal fiber intelligent design model). The formula is as follows:

[0143]

[0144] where n is the number of samples, Y i and respectively represent the actual value (the first-dimensional structural parameters generated in the test set) and the predicted value (the predicted first-dimensional structural parameters). The Adam optimizer updates the weights according to the MSE. The number of iterations is selected as 500 generations. The selected parameters are given in Table 3.

[0145]

[0146] Table 3 Parameters of the DNN Network

[0147] Figure 4 Shows the loss function curves on the training set and test set based on the DNN network according to an embodiment of the present disclosure. In the embodiments of the present disclosure and other possible embodiments, after configuring the parameters of the DNN model network, the DNN network is trained using the training set, and the time taken for training is 11.007 seconds. From Figure 4 it can be seen that the loss function continuously decreases after training until it stabilizes after 400 generations. Finally, after training for 500 generations, the loss function value on the test set is only 0.0669. This also proves that the DNN model can converge when predicting structural parameters, that is, there is no non-uniqueness problem in the dataset, and the method of fixing the working wavelength is feasible.

[0148] Figure 5 Shows the prediction effect of the anti-bending photonic crystal fiber intelligent inverse design model based on the DNN network according to an embodiment of the present disclosure on the test set; where, (a) air hole spacing; (b) the second diameter (small air hole diameter) of the air holes corresponding to the second size smaller than the first size; (c) the first diameter (large air hole diameter) of the air holes corresponding to the first size. As Figure 5 shown, it shows the errors between the predicted values and the true values of the three structural parameters on the test set. The fitting coefficients R 2 are 0.989, 0.888, and 0.922 respectively, and the average fitting coefficient is 0.933. The blue solid line represents the true value, and the red dashed line represents the predicted value. It can be seen that the fitting effect of the DNN for the air hole spacing is better than the first diameter d2 of the air holes corresponding to the first size and the second diameter d1 of the air holes corresponding to the second size. For some values of the first diameter d2 of the air holes corresponding to the first size and the second diameter d1 of the air holes corresponding to the second size, the prediction error of the DNN is relatively obvious. This may be caused by data imbalance, and it may also be that the input data contains too many redundant features, causing the model to ignore important information and pay more attention to invalid information. Therefore, feature screening is required to streamline the input and retain features with strong independence and greater importance.

[0149] In an embodiment of the present disclosure, the method for interpreting each optical fiber bending resistance performance parameter of each dimension of the second-dimension optical fiber bending resistance performance parameters by using a global interpretability algorithm includes: inputting the second-dimension optical fiber bending resistance performance parameters in the intelligent inverse design model of the bending-resistant photonic crystal fiber and the training dataset for training the preset machine learning algorithm or preset network corresponding to the intelligent design of the bending-resistant photonic crystal fiber into a kernel interpreter to construct an interpreter; based on the interpreter, determining the marginal contribution degree of each optical fiber bending resistance performance parameter of each dimension of the second-dimension optical fiber bending resistance performance parameters to the prediction of the first-dimension structural parameters by the intelligent inverse design model of the bending-resistant photonic crystal fiber; and performing weighted averaging on multiple marginal contribution degrees corresponding to all dimensions of optical fiber bending resistance performance parameters to obtain the contribution value corresponding to each optical fiber bending resistance performance parameter of each dimension of the second-dimension optical fiber bending resistance performance parameters.

[0150] In the embodiments of the present disclosure and other possible embodiments, the present disclosure uses a SHAP kernel interpreter (a first interpretability algorithm or a global interpretability algorithm) to perform global interpretability analysis on the prediction of the intelligent design model of the bending-resistant photonic crystal fiber based on a DNN network. First, the intelligent design model of the bending-resistant photonic crystal fiber based on the DNN network is regarded as a cooperative game, where each feature, that is, the 33-dimensional input (multi-dimensional optical fiber bending resistance performance parameters) of the DNN network, is regarded as a cooperative participant; then, the kernel interpreter KernelExplainer of the SHAP library is used to construct an interpreter with the intelligent design model of the bending-resistant photonic crystal fiber based on the DNN network and the training set, so as to obtain the contribution value (importance), that is, the Shapley value, of each feature (each optical fiber bending resistance performance parameter in the multi-dimensional optical fiber bending resistance performance parameters) to the prediction result of the intelligent design model of the bending-resistant photonic crystal fiber based on the DNN network. By calculating the Shapley value of each feature (each optical fiber bending resistance performance parameter in each dimension of the optical fiber bending resistance performance parameters), the contribution of this feature (each optical fiber bending resistance performance parameter in this dimension of the optical fiber bending resistance performance parameters) to the three structural parameters d1, d2, and Λ of the prediction result of the intelligent design model of the bending-resistant photonic crystal fiber based on the DNN network is measured. The parameters d1, d2, and Λ respectively represent the first diameter of the air holes corresponding to the first size in the bending-resistant PCF, the second diameter of the air holes corresponding to the second size smaller than the first size, and the air hole spacing between the air holes.

[0151] Specifically, in the embodiments of the present disclosure and other possible embodiments, an interpreter is constructed based on the intelligent design model of the bend-resistant photonic crystal fiber based on the DNN network (intelligent inverse design model of the bend-resistant photonic crystal fiber) and the training set, so as to obtain the contribution value (importance) of each feature (each fiber bend-resistant performance parameter in each dimension of fiber bend-resistant performance parameters) to the prediction result (multi-dimensional structural parameters) of the intelligent design model of the bend-resistant photonic crystal fiber based on the DNN network (intelligent inverse design model of the bend-resistant photonic crystal fiber). The method includes: inputting the multi-dimensional fiber bend-resistant performance parameters in the intelligent design model of the bend-resistant photonic crystal fiber based on the DNN network (intelligent inverse design model of the bend-resistant photonic crystal fiber) and the training data set for training the DNN network into the kernel interpreter to construct the interpreter; based on the interpreter, determining the marginal contribution degree of each feature (each fiber bend-resistant performance parameter in each dimension of fiber bend-resistant performance parameters) to the prediction result (multi-dimensional structural parameters) of the intelligent design model of the bend-resistant photonic crystal fiber based on the DNN network; performing weighted averaging on the multiple marginal contribution degrees corresponding to the fiber bend-resistant performance parameters of all dimensions (33 dimensions) to obtain the contribution value (importance) corresponding to each fiber bend-resistant performance parameter.

[0152] In the embodiments of the present disclosure and other possible embodiments, this calculation process of performing weighted averaging on the multiple marginal contribution degrees corresponding to the fiber bend-resistant performance parameters of all dimensions to obtain the contribution value (importance) corresponding to each fiber bend-resistant performance parameter involves considering the difference in the structural parameters output by the intelligent design model of the bend-resistant photonic crystal fiber based on the DNN network when the features (all possible subsets corresponding to the multi-dimensional fiber bend-resistant performance parameters) exist and do not exist in all possible subsets of each prediction sample (each structural parameter), so as to obtain the marginal contribution degree of each feature (each fiber bend-resistant performance parameter in each dimension of fiber bend-resistant performance parameters). These marginal contribution degrees of each fiber bend-resistant performance parameter are weighted averaged according to a preset weighting formula (Shapley value formula), and finally the SHAP value (contribution value or importance) representing the independent contribution of this feature (each fiber bend-resistant performance parameter) to the prediction result (multi-dimensional structural parameters) is obtained. By summarizing and visualizing the SHAP values of all features, the SHAP kernel interpreter can show which fiber bend-resistant performance parameters have an important impact on the prediction of the intelligent design model of the bend-resistant photonic crystal fiber based on the DNN network, so as to deeply understand how the intelligent design model of the bend-resistant photonic crystal fiber based on the DNN network makes decisions.

[0153] Among them, in the embodiments of the present disclosure and other possible embodiments, each dimension of fiber bend-resistant performance parameters includes 33 fiber bend-resistant performance parameters, 11 different first bend losses of 11 HOMs, 11 different second bend losses of 11 FMs, and 11 different effective mode areas.

[0154] Figure 6 Shows a summary ranking diagram of the feature importance corresponding to three structural parameters output by the intelligent design model of the bending-resistant photonic crystal fiber based on the DNN network under the SHAP algorithm (under the SHAP algorithm, the overall ranking diagram of the contribution value or importance corresponding to each fiber bending-resistant performance parameter). Figure 6 Gives the three structural parameters d1, d2, and Λ output by the intelligent design model of the bending-resistant photonic crystal fiber based on the DNN network. Under the SHAP algorithm, the summary ranking diagram of the feature importance (the contribution value or importance corresponding to each fiber bending-resistant performance parameter). Among them, class0 / 1 / 2 represent the air hole pitch Λ, the second diameter d1 of the air holes corresponding to the second dimension, and the first diameter d2 of the air holes corresponding to the first dimension, respectively. Among them, Feature 1-11 represent different first bending losses of HOM, Feature 12-22 represent different second bending losses of FM, and Feature 23-33 represent different effective mode areas.

[0155] As can be counted from Figure 6, among the top N (accounting for 1 / 3 of the 33 corresponding to the multi-dimensional fiber bending-resistant performance parameters) more important features (key fiber bending-resistant performance parameters), the effective mode area of the bending-resistant PCF accounts for 6, and the first bending loss and the second bending loss of the bending-resistant PCF account for 5. There are 11 effective mode areas in the test set, and there are 22 first bending losses and second bending losses in total. Therefore, it can be inferred that overall, the effective mode area is more important for the inverse prediction of the structural parameters of the bending-resistant PCF, and the first bending loss and the second bending loss are secondary. This inference also conforms to the true physical meaning, that is, from the positive view, the change of the structural parameters of the bending-resistant PCF can directly affect the change of the effective mode area, and the magnitudes of the first bending loss and the second bending loss are also affected by mode coupling.

[0156] In addition, looking at the three output structural parameters d1, d2, and Λ separately, the ranking of the contribution value or importance corresponding to each fiber bending-resistant performance parameter is slightly different from the overall ranking. Specifically, when predicting the air hole pitch Λ of the bending-resistant PCF, the importance of the effective mode area of the bending-resistant PCF is relatively high compared to the overall ranking of the contribution value or feature importance corresponding to each fiber bending-resistant performance parameter Figure 1However, for the first diameter d2 of the air holes corresponding to the first dimension and the second diameter d1 of the air holes corresponding to the second dimension smaller than the first dimension in the bend-resistant PCF, the importance of the first bend loss and the second bend loss of the bend-resistant PCF is even higher. This is because the size of the effective mode area of the bend-resistant PCF is closely related to the area of the core of the bend-resistant PCF. At the same time, the change in the air hole pitch Λ of the bend-resistant PCF has a more obvious impact on the core of the bend-resistant PCF than the change in the first diameter d2 of the air holes corresponding to the first dimension and the second diameter d1 of the air holes corresponding to the second dimension smaller than the first dimension in the bend-resistant PCF. Therefore, in the reverse prediction task, the effective mode area of the bend-resistant PCF is more important for predicting the air hole pitch Λ of the bend-resistant PCF, and the prediction of the first diameter d2 of the air holes corresponding to the first dimension and the second diameter d2 of the air holes corresponding to the second dimension smaller than the first dimension in the bend-resistant PCF is less important.

[0157] To determine whether the feature importance ranking diagram under the SHAP algorithm is consistent with the physical meaning of the bend-resistant PCF, simulations are carried out in the COMSOL Multiphysics software from the perspective of the physical meaning of the bend-resistant PCF to determine the influence of the structural parameters of the bend-resistant PCF on the performance parameters. Specifically, the first diameter d2 and the second diameter d1 are fixed, and within the set wavelength range, the first bend loss, the second bend loss, and the effective mode area corresponding to different set air hole pitches are obtained.

[0158] Figure 7 Shows the bending performance at different wavelengths and air hole pitches according to an embodiment of the present disclosure; wherein, (a) the first bend loss and the second bend loss; (b) the effective mode area. Figure 7(a) shows the variation of the first bending loss and the second bending loss with wavelength when the pore pitch Λ is set at different values (the different set pore pitches are configured as 24.5 μm, 25 μm, and 25.5 μm). It can be observed that the second bending loss of the FM and the first bending loss of the HOM fluctuate within a certain wavelength range. These fluctuations are due to the coupling between the core mode and the cladding mode. When the wavelength increases from 1.0 μm to 1.4 μm, the fluctuations of the second bending loss of the FM become obvious, while the first bending loss of the HOM continues to increase. In the wavelength range from 1.5 μm to 1.7 μm, the value of the second bending loss of the FM increases with the increase of wavelength, while the first bending loss of the HOM continues to decrease. In the wavelength range from 1.7 μm to 3.0 μm, the value of the second bending loss of the FM generally shows a downward trend, while the value of the first bending loss of the HOM first increases and then decreases. When the air hole pitch is 24.5 μm, the second bending loss of the FM remains below 0.1 dB / m in the wavelength range from 1.1 μm to 2.5 μm, while the first bending loss of the HOM is higher than 1 dB / m, which meets the condition of single-mode operation. Therefore, the bandwidth of single-mode operation is from 1.1 μm to 2.5 μm. When the air hole pitch is 25 μm, the bandwidth of single-mode operation expands to 1.1 μm to 2.8 μm; when the air hole pitch is 25.5 μm, this bandwidth further expands to 1.1 μm to 2.9 μm. Figure 7 (b) shows that as the wavelength increases, the effective mode area gradually increases. It should be noted that when the air hole pitch increases, due to the increase in the core area, the effective mode area also increases accordingly. On the premise of single-mode transportation, the effective mode areas at three different air hole pitches can reach 938.8 μm 2 , 1055.3 μm 2 and 1113.5 μm 2 .

[0159] Figure 8 Shows the bending performance at different wavelengths and large air hole diameters d2 according to an embodiment of the present disclosure; wherein, (a) bending loss; (b) effective mode area. After analyzing the influence of the air hole pitch on single-mode transmission and bending resistance performance, the influence of the air hole diameter is further considered from the inside to the outside. Figure 8 Shows the bending loss and the effective mode area at different wavelengths and the first diameter d2 corresponding to the air holes of the first size. As Figure 8As shown in (a), when the first diameter d2 increases, both the first bending loss of the HOM in the wavelength range of 1.0 - 2.3 μm and the second bending loss of the FM in the wavelength range of 1.0 - 1.7 μm decrease. When the wavelength exceeds 2.5 μm, the influence of the first diameter d2 on the first bending loss and the second bending loss becomes negligible. When the wavelength is 1.5 μm, as d2 increases from 19.6 μm to 20.4 μm, the first bending loss of the HOM decreases from 17.963 dB / m to 9.7782 dB / m, and the second bending loss of the FM decreases from 0.0314 dB / m to 0.02031 dB / m. The increase in d2 is equivalent to the decrease in the hole pitch on both sides of the core, thereby enhancing the confinement ability of the optical fiber and reducing the first bending loss and the second bending loss. For the single-mode bandwidth of the PCF, when d2 is 19.6 μm, its range is between 1.8 - 2.8 μm; when d2 is between 20 - 20.4 μm, this range expands to 1.1 - 2.8 μm. As Figure 8 As shown in (b), it can be observed that the increase in d2 also leads to the decrease of the effective mode area. When the wavelength is 2.8 μm, the effective mode area decreases from 1064.3 μm 2 to 1046.2 μm 2 .

[0160] Figure 9 Shows the bending performance at different wavelengths and small air hole diameters (the second diameter d1) according to the embodiments of the present disclosure; wherein, (a) the first bending loss and the second bending loss; (b) the effective mode area. The variation of the bending loss and the effective mode area with the wavelength and the second diameter d1 corresponding to the air holes of the second dimension. As shown in 9(a), the increase in d1 leads to the decrease of the bending loss. This phenomenon is attributed to the increase in the cladding air hole duty ratio with the increase in d1, which in turn enhances the confinement ability of the optical fiber to light. When d1 is 9.8 μm and 10.2 μm respectively, the second bending loss of the FM at wavelengths of 1.7 μm and 1.9 μm both exceeds 0.1 dB / m, so the range of its single-mode bandwidth decreases. Since the change in the effective mode area mainly depends on the air holes near the core, the influence of d1 on the effective mode area can be ignored, as Figure 9 shown in (b).

[0161] In summary, when the air hole pitch increases, the single-mode operation bandwidth expands, that is, the effective mode area increases, and the first bending loss and the second bending loss show different variations at different wavelengths due to mode coupling; when the large air hole diameter (the first diameter) d2 increases, the bending loss decreases in a specific wavelength range, the single-mode bandwidth expands, and the effective mode area decreases; when the small air hole diameter (the second diameter) d1 increases, the first bending loss and the second bending loss decrease, but the influence on the effective mode area can be ignored, and the single-mode bandwidth range decreases at some wavelengths.

[0162] These physical mechanisms are highly consistent with the SHAP analysis results. The SHAP analysis indicates that overall, the effective mode area is more important for the inverse prediction of structural parameters, followed by the bending loss. From the above physical mechanisms, the change of structural parameters directly affects the effective mode area. For example, when the air hole spacing and the large air hole diameter change, the effective mode area changes accordingly, which is in line with the SHAP analysis. The bending loss is affected by factors such as mode coupling, and its importance is relatively secondary, which also conforms to the SHAP conclusion. Therefore, through the analysis of the physical mechanisms of the influence of structural parameters on performance, the reliability of the SHAP analysis results is further verified.

[0163] In the embodiments of the present disclosure and other possible embodiments, only the top 1 / 3 of the more important features corresponding to the fiber bending resistance performance interpretation parameters in the sorted general diagram are retained (not less than the minimum-dimensional fiber bending resistance performance parameter among the second-dimensional fiber bending resistance performance parameters corresponding to the anti-bending photonic crystal fiber intelligent inverse design optimization model corresponding to the structural parameter fitting coefficient), including: Feature 11, Feature 18, Feature 19, Feature 20, Feature 21, Feature 22, Feature 26, Feature 27, Feature 28, Feature 31, Feature 32, and the remaining 2 / 3 of the second-dimensional fiber bending resistance performance parameters corresponding to the fiber bending resistance performance interpretation parameters are removed, and the minimum-dimensional fiber bending resistance performance parameter among the second-dimensional fiber bending resistance performance parameters corresponding to the anti-bending photonic crystal fiber intelligent inverse design optimization model corresponding to the structural parameter fitting coefficient is determined.

[0164] Figure 10 Show the loss function curves on the training set and validation set based on the DNN network and the second-dimensional fiber bending resistance performance parameters according to the embodiments of the present disclosure. Use the training set corresponding to the second-dimensional fiber bending resistance performance parameters to train the DNN network without changing the network structure; the loss function curves on the training set and validation set based on the DNN network are as Figure 10 shown. It can be Figure 10 seen that the loss function based on the DNN network continuously decreases after training until it stabilizes after 400 generations. Finally, after training for 500 generations, the loss function value on the DNN test set is only 0.0606. The fitting coefficients R2 of the three structural parameters are 0.986, 0.919, and 0.916 respectively, and the average fitting coefficient is 0.940.

[0165] Figure 11 Show the loss function curves on the training set and validation set based on the optimized DNN network and the minimum-dimensional fiber bending resistance performance parameters according to the embodiments of the present disclosure. Use the training set corresponding to the minimum-dimensional fiber bending resistance performance parameters to train the optimized DNN.

[0166] In the embodiments of the present disclosure and other possible embodiments, the network structure is simplified according to the sorting of feature importance to determine the optimized intelligent design network for anti-bending PCF. Specifically, the 33 neurons corresponding to the input layer are reduced to 11 neurons; the middle hidden layer is reduced from 4 layers to 3 layers, and the number of neurons in each layer of the middle hidden layer is configured to be 20; the 3 neurons corresponding to the output layer. Among them, a ReLU (Rectified Linear Unit) activation function is configured after each neuron in the first layer (the first middle hidden layer) of the middle hidden layer. Among them, a Dropout (regularization) layer (regularization function) is configured after each neuron in the second layer (the second middle hidden layer) of the middle hidden layer, and 10% of the corresponding neurons are randomly discarded according to the set dropout rate during the training process (the set dropout rate is configured to be 0.1). Among them, the third layer (the third middle hidden layer) of the middle hidden layer is also configured with 20 neurons, and tanh is configured as the activation function after each neuron in the third layer of the middle hidden layer. Among them, the output value range of tanh is between -1 and 1, which helps to process different data distributions. The last layer is configured as the output layer, including 3 neurons.

[0167] In the embodiments of the present disclosure and other possible embodiments, after reconstructing the corresponding optimized intelligent design network for anti-bending PCF, the optimized intelligent design network for anti-bending PCF based on the DNN network is trained using the training set corresponding to the minimum-dimensional fiber anti-bending performance parameters, and the loss function curves on the training set and the validation set are as Figure 13 shown. It can be seen from the figure that the loss function continuously decreases after training until it stabilizes after 350 generations. Finally, after training for 500 generations, the loss function value on the test set is only 0.054. The fitting coefficients R2 of the three structural parameters are 0.995, 0.907, and 0.930 respectively, and the average fitting coefficient is 0.944. After simplifying the DNN network, the model is optimized in terms of the loss function value and the fitting coefficients of some structural parameters. While maintaining the stability of the overall prediction performance, the prediction accuracy of some parameters is improved. This shows that the simplification operation improves the efficiency and performance of the model to a certain extent, effectively removes redundant information, makes the model structure more reasonable, reduces the computational complexity, and is more conducive to the design of the anti-bending photonic crystal fiber structure in practical applications.

[0168] Embodiments of the present disclosure also propose an intelligent design method for a bend-resistant photonic crystal fiber, which is applied to a forward prediction task and includes: the intelligent design method for a bend-resistant photonic crystal fiber applied to a reverse prediction task as described above; and, based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters, using the intelligent design of the bend-resistant photonic crystal fiber to respectively construct a plurality of intelligent forward design models for predicting the second-dimensional fiber bend-resistant performance parameters by using a corresponding preset machine learning algorithm or a preset network; respectively calculating a plurality of fiber bend-resistant performance fitting coefficients corresponding to the plurality of intelligent forward design models of the bend-resistant photonic crystal fiber; and determining an intelligent optimal forward design model of the bend-resistant photonic crystal fiber based on the plurality of fiber bend-resistant performance fitting coefficients.

[0169] In the forward prediction task, in embodiments of the present disclosure and other possible embodiments, specifically, in the COMSOL Multiphysics physical field simulation software, the finite element method is used to numerically analyze the bending characteristics of a bend-resistant PCF with a cylindrical perfectly matched layer (PML). Among them, when numerically analyzing the bending characteristics of a bend-resistant PCF with a cylindrical PML, the scaling factor and the scaled curvature parameter of the PML are both configured as fixed values (for example, 1). Based on different structural parameters and their set ranges and set change steps corresponding to the air hole spacing, air hole diameter, etc. (fiber structure parameters), under different influencing factors and their set ranges and set change steps corresponding to the operating wavelength, bending radius, etc., data corresponding to the structural parameters of the bend-resistant PCF and the fiber bend-resistant performance parameters corresponding to the structural parameters one by one are generated. A bend-resistant PCF dataset.

[0170] In the forward prediction task, the bend-resistant PCF dataset includes two types of parameters: the structural parameters of the bend-resistant PCF and their corresponding performance parameters. Among them, the intelligent design network of the bend-resistant PCF is configured as the fiber structure parameters (structural parameters) of the bend-resistant PCF, and the output of the machine learning algorithm is configured as the first bending loss of the high-order mode (HOM), the second bending loss of the fundamental mode (FM), and the effective mode area corresponding to the performance parameters.

[0171] In the forward prediction task, when the intelligent design network of the bend-resistant PCF is configured as a machine learning algorithm, the input of the machine learning algorithm is configured as the fiber structure parameters (structural parameters) of the bend-resistant PCF, and the output of the machine learning algorithm is configured as the first bending loss of the high-order mode (HOM), the second bending loss of the fundamental mode (FM), and the effective mode area corresponding to the performance parameters.

[0172] For example, the machine learning algorithm can be configured as an advanced machine learning algorithm such as support vector regression (SVR) or extreme gradient boosting (XGBoost), or the anti-bending PCF intelligent design network configuration can be configured as a deep neural network (DNN).

[0173] In the forward prediction task, among the fiber structure parameters (structural parameters) corresponding to the anti-bending PCF, the first setting range of the air hole pitch Λ is 24.5 - 25.5 μm, and the first setting step is 0.5 μm; the second setting range of the first diameter d2 of the air holes corresponding to the first size is 19.6 - 20.4 μm, and the second setting step is 0.4 μm; the third setting range of the second diameter d1 of the air holes corresponding to the second size is 9.8 - 10.2 μm, and the third setting step is 0.2 μm; the fifth range of the working wavelength λ of the anti-bending PCF is 1.6 - 3.0 μm, and the fifth step is 0.2 μm; the fourth setting range of the bending radius R of the anti-bending PCF is 10 - 16 cm, and the fourth setting step is 2 cm.

[0174] In the forward prediction task, at the corresponding initial setting wavelength of 1.6 μm, fix any three of the pitch or diameter or bending radius corresponding to the air hole pitch Λ, the first diameter d2 of the air holes corresponding to the first size, the second diameter d1 of the air holes corresponding to the second size, and the bending radius R of the anti-bending PCF, and adjust the corresponding pitch or diameter or bending radius within the first setting range of the air hole pitch Λ of 24.5 - 25.5 μm with the first setting step of 0.5 μm or within the second setting range of the first diameter d2 of 19.6 - 20.4 μm with the second setting step of 0.4 μm or within the third setting range of the second diameter d1 of 9.8 - 10.2 μm with the third setting step of 0.2 μm or within the fourth setting range of the bending radius R of 10 - 16 cm with the fourth setting step of 2 cm to obtain the first set of fiber structure parameters (structural parameters) corresponding to the anti-bending PCF and their corresponding first set of performance indicators (performance parameters).

[0175] In the forward prediction task, using the above method, based on the initial setting wavelength of 1.6 μm, with the fifth step, within the fifth range of 1.6 - 3.0 μm, repeat the above experiment to respectively obtain the remaining sets of fiber structure parameters corresponding to the anti-bending PCF within the fifth range of 1.6 - 3.0 μm and their corresponding remaining sets of performance indicators as the forward prediction anti-bending PCF dataset.

[0176] In the forward prediction task, using the COMSOL Multiphysics physical field simulation software, perform numerical analysis on each structural parameter corresponding to the above anti-bending PCF to respectively obtain the first bending loss of the HOM, the second bending loss of the FM, and the effective mode area corresponding to it. Furthermore, a corresponding labeled dataset is generated based on the anti-bending PCF.

[0177] In the forward prediction task, the anti-bending PCF dataset contains 864 groups of data (all groups of fiber structure parameters corresponding to anti-bending PCF and their corresponding performance parameters). For the anti-bending PCF intelligent design network that needs to be trained, the anti-bending PCF dataset is randomly divided into a training set and a test set according to a set ratio of 3:1. The training set and the test set are used to train and verify the model effect respectively.

[0178] In the forward prediction task, considering that the second bending loss of the FM of the anti-bending PCF is too low, most of them are in the order of magnitude of 10 -2 and the difference is small, and the difference in the first bending loss of the HOM of the anti-bending PCF is too large and not smooth enough. Such data characteristics are not conducive to the learning of the anti-bending PCF intelligent design network. Logarithmic processing can compress the range of data into a reasonable interval to obtain the logarithmic first bending loss and the logarithmic second bending loss. Therefore, the first bending loss and the second bending loss corresponding to the HOM and FM of the anti-bending PCF are logarithmically transformed with base 10 to improve the numerical stability of the loss, and the logarithmic first bending loss and the logarithmic second bending loss are obtained.

[0179] In addition, in the forward prediction task, since different variables or features in the anti-bending PCF dataset have different dimensions, and these dimensional differences may cause the anti-bending PCF intelligent design network to overemphasize or ignore certain features during the learning process. Therefore, before training the anti-bending PCF intelligent design network, the training set and the test set need to be standardized. Among them, the standardization method can be configured as the commonly used standardization operations of 0-1 standardization and Z-score standardization.

[0180] In the forward prediction task, the first bending loss and the second bending loss in the corresponding fiber anti-bending performance parameters of the anti-bending PCF dataset are logarithmically transformed (lg) to obtain the logarithmic first bending loss and the logarithmic second bending loss; finally, the structure parameters and the fiber anti-bending performance parameters (including: the logarithmic first bending loss and the logarithmic second bending loss) are pre-standardized to obtain the standardized structure parameters and the standardized fiber anti-bending performance parameters, providing the anti-bending PCF dataset for constructing the subsequent anti-bending PCF intelligent design model training.

[0181] In the forward prediction task, the core of this Z-score standardization method is to divide the original data by the standard deviation after subtracting the mean, so that the mean of the data is 0 and the standard deviation is 1. In order to verify the rationality and necessity of logarithmic transformation and standardization, a comparative experiment was conducted using the SVR model, and the experimental results are shown in Table 2. R 2It represents the fitting situation between the true value and the predicted value. The closer the value is to 1, the more accurate the predicted value is.

[0182] As can be seen from Table 2, the dataset without logarithmic and standardization processing cannot train the intelligent design network of anti-bending PCF to enable it to learn the effective relationship between any structural parameters and the fiber anti-bending performance parameters, and thus cannot complete the corresponding structural parameter prediction. However, only by performing logarithmic operation on the first bending loss and the second bending loss can the prediction effect of the intelligent design model of anti-bending PCF corresponding to SVR be significantly improved. R 2 reaches 0.539. In addition, only performing the standardization operation can also make the intelligent design model of anti-bending PCF converge to a certain extent. The R values after the two standardization operations 2 are 0.601 and 0.707 respectively. From this, it can be seen that for the intelligent design model of anti-bending PCF corresponding to SVR, the Z-score standardization operation is more accurate. Finally, by combining the two preprocessing methods, the fitting coefficient is as high as 0.945, and the corresponding structural parameter prediction is successfully completed. In the subsequent training and prediction of the intelligent design model of anti-bending PCF, logarithmic and standardization preprocessing operations have been performed, so they will not be elaborated one by one.

[0183]

[0184] Comparison results of data preprocessing in Table 2

[0185] Here, five machine learning models are used to complete the forward prediction task (the input is the structural parameter, and the output is the fiber anti-bending performance parameter). The K-nearest neighbor algorithm (KNN) is a lazy, non-parametric algorithm. Its working principle is to find K samples closest to the unknown sample. Generally, a distance metric is used to determine which samples are the closest. After determining the neighboring samples, the output of this sample is jointly determined by the neighboring samples.

[0186] In the forward prediction task, the K in the K-nearest neighbor algorithm used in this disclosure is configured to be 10, the distance metric is selected as the Euclidean distance, and the weight of the neighboring samples is selected as the distance-based weight method, that is, the closer the distance, the higher the weight in the final output. Although KNN does not require training, in actual operation, the samples in the test set look for neighboring samples in the training set. Therefore, in order to improve the accuracy, the ratio of the training set to the test set is re-divided into 9:1. In order to avoid the influence of the K value and the distance metric method on the prediction result, a comparative experiment was carried out. After changing K and the distance metric method, the prediction accuracy of KNN is still relatively low, and the fitting coefficient R between the true value and the predicted value 2 is at most 0.918.

[0187] In the forward prediction task, support vector regression (SVR) is a regression algorithm based on support vector machine (SVM). Among them, SVM is a binary classifier, and its basic idea is to solve a hyperplane that can correctly divide the sample data and has the largest geometric margin. SVR is to solve a hyperplane based on SVM so that the distance from all data points to this plane is the closest, thereby realizing regression.

[0188] In the forward prediction task, a single SVR-based intelligent design model for bend-resistant PCF is only applicable to the prediction of one fiber bend-resistant performance parameter, while the fiber bend-resistant performance parameters corresponding to the output of the above-mentioned bend-resistant PCF include three fiber bend-resistant performance parameters. Therefore, the present disclosure creates a chain of intelligent design models for bend-resistant PCF to use the single-output intelligent design model for bend-resistant PCF in the multi-output intelligent design model for bend-resistant PCF.

[0189] Specifically, the chain of intelligent design models for bend-resistant PCF includes: the first sub-intelligent design model for bend-resistant PCF, the second sub-intelligent design model for bend-resistant PCF, and the third sub-intelligent design model for bend-resistant PCF. Among them, the first sub-intelligent design model for bend-resistant PCF uses the input structural parameters and predicts a corresponding first bending loss for the output; the second sub-intelligent design model for bend-resistant PCF uses the input structural parameters and the output of the first sub-intelligent design model for bend-resistant PCF to predict the corresponding second bending loss; the third sub-intelligent design model for bend-resistant PCF uses the input structural parameters and the first bending loss and the second bending loss corresponding to the outputs of the previous two intelligent design models for bend-resistant PCF to predict the effective mode area. Under the conditions of different hyperparameter penalty factors C and deviation tolerances ε, the effect of the SVR model in predicting the fiber bend-resistant performance parameters is still not good, with a maximum of 0.945.

[0190] In the forward prediction task, the bend-resistant PCF intelligent design model based on SVR has achieved a significant improvement in prediction accuracy compared to the bend-resistant PCF intelligent design model based on KNN, but it still has not reached the optimal.

[0191] Next, in the forward prediction task, the effects of two ensemble algorithms on predicting the bending resistance performance parameters will be studied. An ensemble algorithm is not a single machine learning algorithm, but rather constructs multiple machine learning corresponding intelligent design models for bending-resistant PCFs on a dataset, and takes the fiber bending resistance performance parameters corresponding to all the intelligent design models for bending-resistant PCFs as the final fiber bending resistance performance parameters of this ensemble algorithm according to specific rules. According to the relationship between the base learners, ensemble algorithms can be divided into two major categories: bagging and boosting. The base learners in bagging are generated serially, while the base learners in boosting are generated in parallel. Among them, Random Forest (RF) is a bagging algorithm, and XGBoost is a boosting algorithm. RF mainly improves performance by reducing the model variance, while XGBoost improves performance by reducing the model bias. For the fiber bending resistance performance parameter regression problem of the present disclosure, the base learners of both algorithms are selected as CART trees, and the error function is MSE. The fitting coefficients of the two models are 0.987 and 0.994 respectively, and the training times are 0.104 seconds and 0.076 seconds, indicating that both can efficiently complete the task of predicting the fiber bending resistance performance parameters.

[0192] Finally, in the forward prediction task, an artificial neural network (ANN) with 3 hidden layers, each layer having 50 neurons, was used for the regression task of the present disclosure to predict the fiber bending resistance performance parameters. Figure 12 Shows the prediction effect and training time based on machine learning algorithms according to embodiments of the present disclosure. Figure 12 Summarizes the performance of five machine learning models in terms of prediction accuracy and training time for fiber bending resistance performance parameters. Figure 12 It can be seen that XGBoost has the highest accuracy on the test set and the shortest training time, and RF ranks second. The prediction effects of SVR and KNN are relatively poor. This is because the generalization ability of the ensemble learning model is stronger, and its results inherit the advantages of a large number of base learners. Therefore, the prediction effects of RF and XGBoost are better than those of the single SVR and KNN models. Although ANN performs well on the test set, its training time is the longest, which is more than an order of magnitude longer than the training times of other machine learning algorithms. The training of the ANN model requires continuous iteration to update the weights and bias values in the model using the backpropagation algorithm until the training is completed. Compared with the above four classical machine learning algorithms, the ANN model is more complex and is better at handling more complex tasks. Therefore, the comprehensive performance of ANN is not the best. 2 The training of the ANN model requires continuous iteration to update the weights and bias values in the model using the backpropagation algorithm until the training is completed. Compared with the above four classical machine learning algorithms, the ANN model is more complex and is better at handling more complex tasks. Therefore, the comprehensive performance of ANN is not the best.

[0193] The embodiments of the present disclosure also propose a method for intelligent analysis of bend-resistant photonic crystal fibers, including: applying the intelligent design method as described above to construct an intelligent inverse design optimization model for bend-resistant photonic crystal fibers; and using a local interpretability algorithm to configure each fiber bend-resistant performance parameter corresponding to the minimum fiber bend-resistant performance parameter in the second-dimensional fiber bend-resistant performance parameters in the test set for training the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber as a specific sample; using an explanation instance function to adjust the eigenvalue corresponding to the specific sample to generate a corresponding perturbed sample; based on the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber, the specific sample, the perturbed sample, and the local interpretability algorithm, determining the contribution degree of the local interpretability algorithm to the first-dimensional structural parameter; if the contribution degree of the local interpretability algorithm to the first-dimensional structural parameter is basically consistent with the contribution degree of the global interpretability algorithm to the first-dimensional structural parameter, then the minimum fiber bend-resistant performance parameter corresponding to the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber is optimal.

[0194] In the embodiments of the present disclosure, the method for determining the contribution degree of the local interpretability algorithm to the first-dimensional structural parameter based on the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber, the specific sample, the perturbed sample, and the local interpretability algorithm includes: respectively predicting the first-dimensional structural parameter for the specific sample and the perturbed sample based on the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber to obtain a first predicted structural parameter and a second predicted structural parameter; calculating the similarity between the first predicted structural parameter and the second predicted structural parameter, and configuring the similarity as the weight of the perturbed sample; training the local interpretability algorithm based on the perturbed sample configured with the weight to obtain the contribution degree of each fiber bend-resistant performance parameter corresponding to the minimum fiber bend-resistant performance parameter to the first-dimensional structural parameter.

[0195] In the embodiments of the present disclosure and other possible embodiments, the LIME (Local Interpretable Model-agnostic Explanations) algorithm in the XAI technology introduced by the present invention is an efficient and flexible local interpretation method for analyzing the complex decisions of a deep neural network (DNN) in the design process of bend-resistant PCF. In the design process of bend-resistant PCF, the LIME technology is cleverly applied to the samples in the test set to deeply analyze the prediction behavior of the deep neural network (DNN) model. Among them, the test set includes: the bend-resistant performance parameters of the minimum-dimensional optical fiber and the 3D structure parameters; the bend-resistant performance parameters of the second-dimensional optical fiber in each dimension, including: 11 bend-resistant performance parameters of the minimum-dimensional optical fiber. These samples cover different combinations of design parameters and can comprehensively reflect the performance of the intelligent design model of bend-resistant photonic crystal fiber based on the DNN network under different conditions.

[0196] The LIME algorithm (the second interpretability algorithm or the locality interpretability algorithm) mainly focuses on optimizing the bend-resistant performance parameters of the third-dimensional optical fiber corresponding to the input features of the intelligent design model of bend-resistant photonic crystal fiber based on the DNN network. The bend-resistant performance parameters of the minimum-dimensional optical fiber corresponding to these features are the optical performance indicators we expect, such as the first bending loss of the high-order mode (HOM) of the bend-resistant PCF, the second bending loss of the fundamental mode (FM) of the bend-resistant PCF, and the effective mode area. In the reverse prediction task corresponding to the reverse design model, the performance indicators corresponding to these bend-resistant performance parameters of the minimum-dimensional optical fiber are used as inputs to predict the corresponding structure parameters.

[0197] Specifically, the LIME algorithm selects 11 bend-resistant performance parameters of the minimum-dimensional optical fiber in each dimension of the bend-resistant performance parameters of the second-dimensional optical fiber in the test set as specific samples, and generates corresponding one-dimensional adjacent perturbation samples based on these specific samples. Among them, the method of generating one-dimensional groups of adjacent perturbation samples based on these specific samples includes: using the interpretation instance function to adjust the feature values corresponding to the specific samples based on the 11 bend-resistant performance parameters of the minimum-dimensional optical fiber corresponding to the bend-resistant performance parameters of the second-dimensional optical fiber in each dimension, and generating the corresponding one-dimensional adjacent perturbation samples.

[0198] Among them, the interpretation instance function for generating the corresponding one-dimensional adjacent perturbation samples configures the explain_instance function of the lime.lime_tabular.LimeTabularExplainer class in the LIME algorithm.

[0199] These perturbed samples are obtained by slightly adjusting the eigenvalues of specific samples, aiming to cover the local regions of the feature space corresponding to the 11 smallest fiber bending resistance performance parameters of the second dimension in each dimension of the test set. Then, the pre-trained intelligent design model of bending-resistant PCF based on the DNN network is used to predict the structural parameters of these perturbed samples and specific samples respectively, obtaining the (first dimension) first predicted structural parameter and the (first dimension) second predicted structural parameter. Subsequently, the LIME algorithm adopts a weighted linear regression model as a local interpreter. This interpreter assigns different weights according to the distance between the perturbed samples and the specific samples, calculates the similarity between the first predicted structural parameter corresponding to the perturbed samples and the second predicted structural parameter corresponding to the specific samples, and the similarity is configured to assign weights to the perturbed samples, so that the specific samples close to the original instance have a higher influence in the local linear model, thereby capturing the non-linear response pattern of the local linear model in the locally sensitive region. The LIME algorithm is trained with the weighted perturbed samples to approximate the prediction of the structural parameters by the intelligent design model of bending-resistant PCF based on the DNN network in the local region, and the contribution degree or importance of each feature of the 11 optimized third dimension fiber bending resistance performance parameters of the second dimension fiber bending resistance performance parameter in each dimension to the structural parameter prediction result is obtained.

[0200] Through this process, the LIME algorithm can construct a local linear model, which can intuitively show the contribution degree or importance of the 11 smallest fiber bending resistance performance parameters of the second dimension in each dimension to the structural parameter prediction. Finally, the LIME algorithm can identify which structural parameters have the most critical impact on the performance of bending-resistant PCF under specific design conditions. This local interpretation not only enhances the transparency of the intelligent design model of bending-resistant PCF based on the DNN network and improves the trust of designers, but also provides a scientific basis for further optimizing design parameters and improving fiber performance. Through the application of LIME, we can understand the decision-making process of the DNN model more deeply, thus promoting the continuous development and progress of the technology of bending-resistant photonic crystal fibers.

[0201] Figure 13The figure shows the impact analysis diagram of three structural parameters under the LIME algorithm according to an embodiment of the present disclosure. Under the LIME algorithm, for the impact analysis diagram of three structural parameters, it can be seen that for the air hole pitch Λ in the bending-resistant PCF, the input feature X31 represents the effective mode area of the FM, showing a positive number 0.16, that is, the effective mode area of the FM has the largest positive contribution to the air hole pitch Λ; for the second diameter d1 of the air holes corresponding to the second size smaller than the first size in the bending-resistant PCF, the input feature X21 represents the second bending loss of the FM, showing a positive number 0.38, that is, the second bending loss of the FM has the largest positive contribution to the second diameter d1. At the same time, the input features X6, X8, and X10 all represent the second bending loss of the HOM and also have a positive contribution to the second diameter d1; for the first diameter d2 of the air holes corresponding to the first size in the bending-resistant PCF, the input feature X32 represents the effective mode area of the FM, showing a positive number 0.73, that is, the effective mode area of the FM has the largest positive contribution to the first diameter d2. At the same time, the input features X11, X14, X15, and X16 all represent the second bending loss of the FM and also play an important role in the first diameter d2. This is consistent with the results obtained by SHAP in the previous text. From the perspective of the physical mechanism, the effective mode area is closely related to the core area, and the change of the air hole pitch has an obvious impact on the core. Therefore, the effective mode area of the FM is important for predicting the air hole pitch. The bending loss is affected by factors such as mode coupling. When predicting the air hole diameters corresponding to the first size and the second size, the importance of the bending loss is reflected, which is also consistent with the LIME analysis results. Among them, X represents the above-mentioned Feature.

[0202] Through the analysis of two interpretive machine learning algorithms, SHAP and LIME, we have a deeper understanding of the structural parameters of the bending-resistant photonic crystal fiber. The results of the LIME algorithm reveal the specific contributions of each input feature to the structural parameters, and the differences in the feature importance rankings also provide us with valuable information for fiber design and optimization.

[0203] In summary, the intelligent design, analysis method and system of the bend-resistant photonic crystal fiber of the present invention significantly improve the design efficiency and accuracy of the intelligent design model of the bend-resistant performance of the photonic crystal fiber by combining the finite element analysis method, machine learning algorithm and interpretable artificial intelligence technology. By constructing a forward prediction model and a reverse design model, efficient prediction from structural parameters to performance indicators and precise design from performance indicators to structural parameters are achieved. At the same time, using interpretable tools such as SHAP and LIME, the model decision-making process is deeply analyzed, and the key features affecting the design parameters of the bend-resistant photonic crystal fiber are identified, providing a scientific basis for optimizing the fiber design. The implementation of the present invention not only overcomes the limitations of the traditional design method of bend-resistant PCF, shortens the design cycle of the intelligent design model of bend-resistant PCF, reduces the cost of the intelligent design model of bend-resistant PCF, but also lays a solid foundation for the wide application of photonic crystal fiber in the fields of optical communication, sensing, etc. In the future, with the continuous development of artificial intelligence technology, this method will be further optimized to explore the design of more high-performance photonic crystal fibers and promote the continuous progress of fiber technology.

[0204] The present invention provides an intelligent design, analysis method and system for a bend-resistant photonic crystal fiber (PCF), aiming to solve the problems in the traditional design method of bend-resistant PCF that the calculation resources of structural parameters consume a large amount, resulting in a long design cycle of bend-resistant PCF, and there are serious deficiencies in the interpretability of the structural parameters of the bend-resistant PCF and the transparency of the design process of the intelligent design model of bend-resistant PCF. By introducing the technology of Explainable Artificial Intelligence (XAI), with its powerful feature analysis ability, the internal relationship between the structural parameters of the bend-resistant PCF design and the fiber bend-resistant performance parameters of the bend-resistant PCF is accurately revealed, and the fiber bend-resistant performance parameters affecting the structural parameters corresponding to the bend-resistant PCF (photonic crystal fiber) design are identified, thereby screening out redundant fiber bend-resistant performance parameters, opening up a new path for the design of bend-resistant PCF, further reducing the complexity of the intelligent design model of bend-resistant PCF, and improving the calculation efficiency of the intelligent design model of bend-resistant PCF, strongly promoting its wide application in the fields of scientific research and engineering.

[0205] The present invention provides an intelligent design method for a bend-resistant photonic crystal fiber, comprising the following steps: 1. Construction of a bend-resistant PCF dataset corresponding to the bend-resistant PCF: Based on the structure of the bend-resistant PCF, using a physical field simulation software (COMSOL Multiphysics software), data corresponding to the structural parameters of the bend-resistant PCF and the fiber bend-resistant performance parameters corresponding one-to-one to the structural parameters are generated. Then, the first bend loss and the second bend loss in the fiber bend-resistant performance parameters corresponding to the bend-resistant PCF dataset are logarithmized (lg) to obtain the logarithmized first bend loss and the logarithmized second bend loss; finally, the structural parameters and the fiber bend-resistant performance parameters (including: the logarithmized first bend loss and the logarithmized second bend loss) are preprocessed by standardization to obtain the standardized structural parameters and the standardized fiber bend-resistant performance parameters, providing a bend-resistant PCF dataset for constructing the subsequent intelligent design model training for the bend-resistant PCF. The core task of constructing the bend-resistant PCF dataset corresponding to the bend-resistant PCF is to construct a comprehensive and accurate bend-resistant PCF dataset. Based on the structural parameters such as the air hole spacing and the air hole diameter (fiber structural parameters) corresponding to different structural parameters and their set ranges and set change steps, and under different influencing factors such as the working wavelength and the bending radius corresponding to their set ranges and set change steps, using the physical field simulation software, data corresponding to the structural parameters of the bend-resistant PCF and the fiber bend-resistant performance parameters corresponding one-to-one to the structural parameters are generated. 2. Training of the bend-resistant PCF intelligent design model: Using the above standardized bend-resistant PCF dataset to train the bend-resistant PCF intelligent design network, enabling the bend-resistant PCF intelligent design network to learn the mapping relationship between the fiber bend-resistant performance index (fiber bend-resistant performance parameters) and the structural parameters, and obtaining the corresponding bend-resistant PCF intelligent design model; thus, based on the bend-resistant PCF intelligent design model, using the second-dimensional fiber bend-resistant performance parameters, the corresponding first-dimensional structural parameters for predicting the bend-resistant PCF are obtained. The bend-resistant PCF intelligent design network can be configured to support advanced machine learning algorithms or networks such as support vector regression (SVR), extreme gradient boosting (XGBoost), and deep neural network (DNN); with the aid of the bend-resistant PCF dataset corresponding to the standardized structural parameters and the standardized fiber bend-resistant performance parameters, the bend-resistant PCF intelligent design network is trained and fine-tuned.3. Optimization of the intelligent design model for bend-resistant PCF: Use XAI technology to perform global interpretability analysis based on the SHAP (Shapley Additive Explanations) algorithm and local interpretability analysis based on the LIME (Local Interpretable Model-agnostic Explanations) algorithm on the intelligent design model for bend-resistant PCF. Based on the importance of the second-dimensional fiber bend-resistant performance parameters corresponding to the global interpretability analysis and / or local interpretability analysis, optimize and adjust the second-dimensional fiber bend-resistant performance parameters corresponding to the intelligent design model for bend-resistant PCF, eliminate unimportant second-dimensional fiber bend-resistant performance parameters, and obtain optimized third-dimensional fiber bend-resistant performance parameters (updated-dimensional fiber bend-resistant performance parameters or latest updated-dimensional fiber bend-resistant performance parameters or minimum-dimensional fiber bend-resistant performance parameters). Use the optimized third-dimensional fiber bend-resistant performance parameters (updated-dimensional fiber bend-resistant performance parameters or latest updated-dimensional fiber bend-resistant performance parameters or minimum-dimensional fiber bend-resistant performance parameters) and their corresponding structural parameters to retrain the optimized intelligent design network for bend-resistant PCF, thereby improving the prediction accuracy and generalization ability of the intelligent design model for bend-resistant PCF for structural parameters.

[0206] Through the collaborative operation of the SHAP algorithm and the LIME algorithm, the intelligent design model for bend-resistant PCF is comprehensively and deeply analyzed from two dimensions corresponding to the global fiber bend-resistant performance interpretation parameters corresponding to global interpretability and / or local fiber bend-resistant performance interpretation parameters.

[0207] Specifically, based on the second-dimensional fiber anti-bending performance parameters corresponding to the intelligent design model of anti-bending PCF, a global interpretability analysis is carried out to determine the contribution degree or importance of the global fiber anti-bending performance interpretation parameters; sort the global fiber anti-bending performance interpretation parameters according to the contribution degree or importance from large to small to obtain the sorted second-dimensional fiber anti-bending performance parameters; take the set of the first N sorted global fiber anti-bending performance parameters from the sorted second-dimensional fiber anti-bending performance parameters to obtain the optimized third-dimensional fiber anti-bending performance parameters (updated-dimensional fiber anti-bending performance parameters or latest updated-dimensional fiber anti-bending performance parameters or minimum-dimensional fiber anti-bending performance parameters); adjust the input layer of the anti-bending PCF intelligent design network based on the optimized third-dimensional fiber anti-bending performance parameters (updated-dimensional fiber anti-bending performance parameters or latest updated-dimensional fiber anti-bending performance parameters or minimum-dimensional fiber anti-bending performance parameters) so that the number of the input layer of the anti-bending PCF intelligent design network matches the number corresponding to the optimized third-dimensional fiber anti-bending performance parameters (updated-dimensional fiber anti-bending performance parameters or latest updated-dimensional fiber anti-bending performance parameters or minimum-dimensional fiber anti-bending performance parameters) to obtain the optimized anti-bending PCF intelligent design network; use the optimized third-dimensional fiber anti-bending performance parameters (updated-dimensional fiber anti-bending performance parameters or latest updated-dimensional fiber anti-bending performance parameters or minimum-dimensional fiber anti-bending performance parameters) to retrain the optimized anti-bending PCF intelligent design network, effectively improving the prediction accuracy and generalization ability of the model.

[0208] With the powerful learning and mapping capabilities of the optimized anti-bending PCF intelligent design model, based on the input optimized third-dimensional fiber anti-bending performance parameters (updated-dimensional fiber anti-bending performance parameters or latest updated-dimensional fiber anti-bending performance parameters or minimum-dimensional fiber anti-bending performance parameters), quickly predict the structural parameters of the anti-bending PCF that meet the requirements of the optimized third-dimensional fiber anti-bending performance parameters (updated-dimensional fiber anti-bending performance parameters or latest updated-dimensional fiber anti-bending performance parameters or minimum-dimensional fiber anti-bending performance parameters), covering key parameters such as air hole spacing and air hole diameters of different sizes.

[0209] The execution entity of the intelligent design and analysis method of the anti-bending photonic crystal fiber can be an image processing device. For example, the intelligent design and analysis method of the anti-bending photonic crystal fiber can be executed by a terminal device or a server or other processing devices. Among them, the terminal device can be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementation manners, the intelligent design and analysis method of the anti-bending photonic crystal fiber can be implemented by a processor invoking computer-readable instructions stored in a memory.

[0210] According to one aspect of the present disclosure, there is provided an intelligent design system for a bend-resistant photonic crystal fiber, which is applied to a reverse prediction task, and includes: a first generation unit configured to generate first-dimensional structural parameters of the bend-resistant photonic crystal fiber and second-dimensional fiber bend-resistant performance parameters corresponding to and greater than the first-dimensional structural parameters; a first construction unit configured to construct an intelligent reverse design model of the bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters based on the first-dimensional structural parameters and the corresponding second-dimensional fiber bend-resistant performance parameters; a first calculation unit configured to calculate a structural parameter fitting coefficient corresponding to the intelligent reverse design model of the bend-resistant photonic crystal fiber; and a second construction unit configured to construct an intelligent reverse design optimization model of the bend-resistant photonic crystal fiber not lower than the structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter among the corresponding second-dimensional fiber bend-resistant performance parameters.

[0211] According to one aspect of the present disclosure, there is provided an intelligent design system for a bend-resistant photonic crystal fiber, which is applied to a reverse prediction task, and includes: an electronic device configured with a processor and a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the above intelligent design method.

[0212] According to one aspect of the present disclosure, there is provided an intelligent design system for a bend-resistant photonic crystal fiber, which is applied to a reverse prediction task, and includes: a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the above intelligent design method is implemented.

[0213] According to one aspect of the present disclosure, there is provided an intelligent design system for a bend-resistant photonic crystal fiber, which is applied to a reverse prediction task, and includes: a computer program product having a computer program / instructions set therein, and when the computer program / instructions are executed by a processor, the intelligent design method as described above is implemented.

[0214] According to one aspect of the present disclosure, there is provided an intelligent design of a bend-resistant photonic crystal fiber for a forward prediction task, including: the intelligent design system of the bend-resistant photonic crystal fiber for a reverse prediction task as described above; and a third construction unit configured to, based on the first-dimensional structure parameters and their corresponding second-dimensional fiber bend-resistant performance parameters, use a preset machine learning algorithm or a preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber to construct a plurality of intelligent forward design models of the bend-resistant photonic crystal fiber for predicting the second-dimensional fiber bend-resistant performance parameters; a second calculation unit configured to calculate respectively a plurality of fiber bend-resistant performance fitting coefficients corresponding to the plurality of intelligent forward design models of the bend-resistant photonic crystal fiber; and a determination unit configured to determine an intelligent optimal forward design model of the bend-resistant photonic crystal fiber based on the plurality of fiber bend-resistant performance fitting coefficients.

[0215] According to one aspect of the present disclosure, there is provided an intelligent design of a bend-resistant photonic crystal fiber for a forward prediction task, including: an electronic device configured with a processor and a memory for storing instructions executable by the processor; wherein the processor is configured to call the instructions stored in the memory to execute the intelligent design method as described above.

[0216] According to one aspect of the present disclosure, there is provided an intelligent design of a bend-resistant photonic crystal fiber for a forward prediction task, including: a computer-readable storage medium having computer program instructions stored thereon, and the computer program instructions, when executed by a processor, implement the intelligent design method as described above.

[0217] According to one aspect of the present disclosure, there is provided an intelligent design of a bend-resistant photonic crystal fiber for a forward prediction task, including: a computer program product having a computer program / instructions set therein, and the computer program / instructions, when executed by a processor, implement the intelligent design method as described above.

[0218] According to one aspect of the present disclosure, an intelligent analysis system for a bend-resistant photonic crystal fiber is provided, including: an intelligent design system for a bend-resistant photonic crystal fiber applied to the reverse prediction task as described above; and a locality interpretability unit for using a locality interpretability algorithm to configure each dimension of the fiber bend-resistant performance parameter corresponding to the minimum dimension of the fiber bend-resistant performance parameter in the test set for training the intelligent reverse design optimization model of the bend-resistant photonic crystal fiber as a specific sample; a perturbation sample generation unit for using an interpretation instance function to adjust the eigenvalue corresponding to the specific sample to generate a corresponding perturbation sample; a contribution degree determination unit for determining the contribution degree to the first dimension of structural parameters determined by the locality interpretability algorithm based on the intelligent reverse design optimization model of the bend-resistant photonic crystal fiber, the specific sample, the perturbation sample, and the locality interpretability algorithm; a judgment unit for determining that the minimum dimension of the fiber bend-resistant performance parameter corresponding to the intelligent reverse design optimization model of the bend-resistant photonic crystal fiber is optimal if the contribution degree to the first dimension of structural parameters determined by the locality interpretability algorithm is basically consistent with the contribution degree to the first dimension of structural parameters determined by the global interpretability algorithm.

[0219] According to one aspect of the present disclosure, an intelligent analysis system for a bend-resistant photonic crystal fiber is provided, including: an electronic device configured with a processor and a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the above intelligent analysis method.

[0220] According to one aspect of the present disclosure, an intelligent analysis system for a bend-resistant photonic crystal fiber is provided, including: a computer-readable storage medium storing computer program instructions thereon, and when the computer program instructions are executed by a processor, the intelligent analysis method as described above is implemented.

[0221] According to one aspect of the present disclosure, an intelligent analysis system for a bend-resistant photonic crystal fiber is provided, including: a computer-readable storage medium storing computer program instructions thereon, and when the computer program instructions are executed by a processor, the intelligent analysis method as described above is implemented.

[0222] Those skilled in the art can understand that in the intelligent design and analysis methods of the bend-resistant photonic crystal fiber in the above specific embodiments, the writing order of each step does not mean a strict execution order and does not constitute any limitation to the implementation process. The specific execution order of each step should be determined according to its function and possible internal logic.

[0223] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the methods described in the above embodiments of the intelligent design and analysis method of the anti-bending photonic crystal fiber. The specific implementation can refer to the description of the above method embodiments. For the sake of brevity, it will not be repeated here.

[0224] The embodiments of the present disclosure also propose a computer-readable storage medium, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the intelligent design and analysis method of the above anti-bending photonic crystal fiber is implemented. The computer-readable storage medium can be a non-volatile computer-readable storage medium.

[0225] The present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for causing a processor to implement various aspects of the present disclosure are loaded.

[0226] The computer-readable storage medium can be a tangible device that can hold and store instructions used by an instruction execution device. The computer-readable storage medium can be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. More specific examples (non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punched card or raised structures in a groove storing instructions thereon, and any suitable combination of the above. The computer-readable storage medium used herein is not construed as an instantaneous signal itself, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated through a waveguide or other transmission medium (e.g., an optical pulse through an optical fiber cable), or an electrical signal transmitted through a wire.

[0227] The embodiments of the present disclosure also propose an electronic device, including: a processor; a memory for storing processor-executable instructions; wherein, the processor is configured to perform the intelligent design and analysis method of the above anti-bending photonic crystal fiber. Among them, the electronic device can be provided as a terminal, a server, or other forms of devices.

[0228] In an exemplary embodiment, the electronic device may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components for performing the above method.

[0229] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, when executed by the processor of the computer or other programmable data processing apparatus, create a means for implementing the functions / acts specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions may also be stored in a computer-readable storage medium that, when read and executed, causes a computer, a programmable data processing apparatus, and / or other devices to operate in a particular manner, such that the computer-readable medium storing the instructions comprises a manufacture including instructions which implement various aspects of the functions / acts specified in one or more blocks of the flowchart and / or block diagram.

[0230] The computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in one or more blocks of the flowchart and / or block diagram.

[0231] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagram may represent a module, a segment of code, or a portion of an instruction, which comprises one or more executable instructions for implementing the specified logical function. In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two consecutive blocks may in fact be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending upon the functionality involved. It should also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations of blocks in the block diagrams and / or flowchart illustrations, can be implemented by special-purpose hardware-based systems that perform the specified functions or acts, or combinations of special-purpose hardware and computer instructions.

[0232] The embodiments of the present disclosure have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of the technology in the market, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. An intelligent design method for a bend-resistant photonic crystal fiber, applied to reverse prediction tasks, characterized in that, Including: Generating the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the second-dimensional fiber bend-resistant performance parameters of the dimension greater than the first-dimensional structural parameters; based on the first-dimensional structural parameters and the corresponding second-dimensional fiber bend-resistant performance parameters, constructing an intelligent inverse design model of the bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters; Calculating the structural parameter fitting coefficient corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber; Constructing an intelligent inverse design optimization model of the bend-resistant photonic crystal fiber not less than the structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter among the corresponding second-dimensional fiber bend-resistant performance parameters.

2. The intelligent design method according to claim 1, characterized in that, The method for generating the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the second-dimensional fiber bend-resistant performance parameters of the dimension greater than the first-dimensional structural parameters includes: using physical field simulation software, under the set working wavelength and the first-dimensional structural parameters corresponding to the bend-resistant photonic crystal fiber, adjusting the bending radius corresponding to the photonic crystal fiber to generate the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the second-dimensional fiber bend-resistant performance parameters of the dimension greater than the first-dimensional structural parameters; and / or, The method for adjusting the bending radius corresponding to the photonic crystal fiber under the set working wavelength and the first-dimensional structural parameters corresponding to the bend-resistant photonic crystal fiber to generate the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the corresponding second-dimensional fiber bend-resistant performance parameters includes: Step S201: Configuring the first-dimensional structural parameters as the first diameter of the air holes corresponding to the first size in the bend-resistant photonic crystal fiber, the second diameter of the air holes corresponding to the second size smaller than the first size, and the air hole spacing between the air holes; Step S202: Obtaining the set range and set step length corresponding to the first diameter, the second diameter, the air hole spacing, and the bending radius; Step S203: Configuring any one of the structural parameters among the first diameter, the second diameter, and the air hole spacing as a variable parameter, and configuring the structural parameters other than the variable parameter as constant parameters; Step S204: Starting from any endpoint of the set range corresponding to the variable parameter, adjusting the bending radius corresponding to the photonic crystal fiber under the set working wavelength, and determining the parameter change corresponding to the variable parameter according to the corresponding set step; Step S205: Determining the corresponding second-dimensional fiber bend-resistant performance parameters under the parameter change corresponding to the variable parameter and the constant parameters; Looping or repeating the above steps S203-S205 until the first diameter, the second diameter, and the air hole spacing are configured as variable parameters, thereby generating the first-dimensional structural parameters of the bend-resistant photonic crystal fiber and the corresponding second-dimensional fiber bend-resistant performance parameters; and / or, Before constructing an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters, logarithmic processing is performed on the first bending loss of the high-order mode and the second bending loss of the fundamental mode in the second-dimensional fiber bend-resistant performance parameters to obtain the logarithmized first bending loss and the logarithmized second bending loss; standardization processing is performed on the logarithmized first bending loss, the logarithmized second bending loss, and the effective mode area in the second-dimensional fiber bend-resistant performance parameters.

3. The intelligent design method according to any one of claims 1 or 2, characterized in that The method for constructing an intelligent inverse design model of a bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters includes: determining a training set based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters; using the training set to train a preset machine learning algorithm or a preset network corresponding to the intelligent design of the bend-resistant photonic crystal fiber to construct an intelligent inverse design model of the bend-resistant photonic crystal fiber for predicting the first-dimensional structural parameters; and / or, The method for calculating the structural parameter fitting coefficient corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber includes: determining a test set based on the first-dimensional structural parameters and their corresponding second-dimensional fiber bend-resistant performance parameters; using the test set to calculate the structural parameter fitting coefficient corresponding to the intelligent inverse design model of the bend-resistant photonic crystal fiber.

4. The intelligent design method according to any one of claims 1-3, characterized in that, The method for constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber not lower than the corresponding structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter in the corresponding second-dimensional fiber bend-resistant performance parameters includes: using a global interpretability algorithm to interpret the second-dimensional fiber bend-resistant performance parameters, each fiber bend-resistant performance parameter in the second-dimensional fiber bend-resistant performance parameters; sorting the second-dimensional fiber bend-resistant performance parameters based on the contribution degree corresponding to each fiber bend-resistant performance parameter to obtain the sorted second-dimensional fiber bend-resistant performance parameters; selecting the corresponding updated-dimensional fiber bend-resistant performance parameters from the sorted second-dimensional fiber bend-resistant performance parameters according to the contribution degree corresponding to each fiber bend-resistant performance parameter and a set contribution degree; constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber not lower than the corresponding structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter in the corresponding second-dimensional fiber bend-resistant performance parameters based on the updated-dimensional fiber bend-resistant performance parameters, the first-dimensional structural parameters, and the corresponding optimized preset machine learning algorithm or preset network; and / or, The method for constructing an intelligent inverse design optimization model of a bend-resistant photonic crystal fiber not lower than the corresponding structural parameter fitting coefficient and the minimum-dimensional fiber bend-resistant performance parameter in the corresponding second-dimensional fiber bend-resistant performance parameters based on the updated-dimensional fiber bend-resistant performance parameters, the first-dimensional structural parameters, and the corresponding optimized preset machine learning algorithm or preset network includes: Step 301: Optimize the corresponding preset machine learning algorithm or preset network for the intelligent design of the bend-resistant photonic crystal fiber by using the updated-dimensional fiber bend-resistant performance parameters; Step 302: Train the optimized preset machine learning algorithm or preset network by using the updated-dimensional fiber bend-resistant performance parameters and their corresponding first-dimensional structure parameters; Step 303: Calculate the fitting coefficient of the optimized structure parameters of the intelligent inverse design optimization model of the bend-resistant photonic crystal fiber corresponding to the optimized preset machine learning algorithm or preset network; Step 304: Compare the fitting coefficient of the optimized structure parameters with the fitting coefficient of the structure parameters corresponding to the second-dimensional fiber bend-resistant performance parameters; Step 304: If the fitting coefficient of the optimized structure parameters is greater than or equal to the fitting coefficient of the structure parameters corresponding to the second-dimensional fiber bend-resistant performance parameters, obtain the set deletion dimension; delete the fiber bend-resistant performance parameters corresponding to the set deletion dimension with a lower ranking from the updated-dimensional fiber bend-resistant performance parameters to obtain the latest updated-dimensional fiber bend-resistant performance parameters; Based on the latest updated-dimensional fiber bend-resistant performance parameters; repeat Steps 301 to 304 until an intelligent inverse design optimization model of the bend-resistant photonic crystal fiber not lower than the minimum-dimensional fiber bend-resistant performance parameter in the second-dimensional fiber bend-resistant performance parameters corresponding to the fitting coefficient of the structure parameters is constructed; Step 305: If the fitting coefficient of the optimized structure parameters is less than the fitting coefficient of the structure parameters corresponding to the second-dimensional fiber bend-resistant performance parameters, obtain the set addition dimension; add the fiber bend-resistant performance parameters corresponding to the set deletion dimension with a higher ranking from the updated-dimensional fiber bend-resistant performance parameters to obtain the latest updated-dimensional fiber bend-resistant performance parameters; Based on the latest updated-dimensional fiber bend-resistant performance parameters, repeat Steps 301 to 303, Step 305 until an intelligent inverse design optimization model of the bend-resistant photonic crystal fiber not lower than the minimum-dimensional fiber bend-resistant performance parameter in the second-dimensional fiber bend-resistant performance parameters corresponding to the fitting coefficient of the structure parameters is constructed.

5. The intelligent design method according to claim 4, wherein The method for optimizing the corresponding preset machine learning algorithm or preset network for the intelligent design of the bend-resistant photonic crystal fiber by using the updated-dimensional fiber bend-resistant performance parameters includes: When selecting the corresponding updated-dimensional fiber bending resistance performance parameter from the second-ranked fiber bending resistance performance parameters or deleting the fiber bending resistance performance parameter corresponding to the set deletion dimension with a lower ranking from the updated-dimensional fiber bending resistance performance parameters to obtain the latest updated-dimensional fiber bending resistance performance parameter, the number of neurons corresponding to the input layer of the preset machine learning algorithm or preset network corresponding to the intelligent design of the bending-resistant photonic crystal fiber is adjusted by using the updated-dimensional fiber bending resistance performance parameter or the latest updated-dimensional fiber bending resistance performance parameter; according to the set layer step size, the number of layers of the intermediate hidden layer and the number of neurons in each layer of the preset machine learning algorithm or preset network corresponding to the intelligent design of the bending-resistant photonic crystal fiber are reduced to obtain an optimized preset machine learning algorithm or preset network; wherein, the number of neurons corresponding to the input layer is the same as the dimension corresponding to the updated-dimensional fiber bending resistance performance parameter or the latest updated-dimensional fiber bending resistance performance parameter; or, when adding the fiber bending resistance performance parameter corresponding to the set deletion dimension with a higher ranking from the updated-dimensional fiber bending resistance performance parameters to obtain the latest updated-dimensional fiber bending resistance performance parameter, the number of neurons corresponding to the input layer of the preset machine learning algorithm or preset network corresponding to the intelligent design of the bending-resistant photonic crystal fiber is adjusted by using the updated-dimensional fiber bending resistance performance parameter; according to the set layer step size, the number of layers of the intermediate hidden layer and the number of neurons in each layer of the preset machine learning algorithm or preset network corresponding to the intelligent design of the bending-resistant photonic crystal fiber are increased to obtain an optimized preset machine learning algorithm or preset network; wherein, the number of neurons corresponding to the input layer is the same as the dimension corresponding to the latest updated-dimensional fiber bending resistance performance parameter; and / or, The method for interpreting each fiber bending resistance performance parameter in the second-dimensional fiber bending resistance performance parameter by using the global interpretability algorithm includes: inputting the second-dimensional fiber bending resistance performance parameter in the intelligent inverse design model of the bending-resistant photonic crystal fiber and the training dataset for training the preset machine learning algorithm or preset network corresponding to the intelligent design of the bending-resistant photonic crystal fiber into the kernel interpreter to construct an interpreter; based on the interpreter, determining the marginal contribution degree of each fiber bending resistance performance parameter in the second-dimensional fiber bending resistance performance parameter to predicting the first-dimensional structural parameter by the intelligent inverse design model of the bending-resistant photonic crystal fiber; and performing weighted average on multiple marginal contribution degrees corresponding to all-dimensional fiber bending resistance performance parameters to obtain the contribution value corresponding to each fiber bending resistance performance parameter in the second-dimensional fiber bending resistance performance parameter.

6. An intelligent design method for a bend-resistant photonic crystal fiber, which is applied to a forward prediction task, is characterized in that, Including: The intelligent design method of the bending-resistant photonic crystal fiber according to any one of claims 1-5 applied to the inverse prediction task; And, Based on the first - dimensional structural parameters and their corresponding second - dimensional fiber bending resistance performance parameters of the anti - bending photonic crystal fiber, using the corresponding preset machine learning algorithm or preset network, respectively construct multiple intelligent forward design models of anti - bending photonic crystal fiber for predicting the second - dimensional fiber bending resistance performance parameters; respectively calculate multiple fiber bending resistance performance fitting coefficients corresponding to the multiple intelligent forward design models of anti - bending photonic crystal fiber; Based on the multiple fiber bending resistance performance fitting coefficients, determine the intelligent optimal forward design model of anti - bending photonic crystal fiber.

7. A method for intelligent analysis of a bend-resistant photonic crystal fiber, characterized in that, Including: Apply the intelligent design method according to any one of claims 1 - 6 to construct an intelligent reverse design optimization model of anti - bending photonic crystal fiber; and, Using the local interpretability algorithm, configure each - dimensional fiber bending resistance performance parameter corresponding to the minimum - dimensional fiber bending resistance performance parameter in the test set for training the intelligent reverse design optimization model of anti - bending photonic crystal fiber as a specific sample; use the interpretation instance function to adjust the eigenvalue corresponding to the specific sample to generate a corresponding perturbed sample; based on the intelligent reverse design optimization model of anti - bending photonic crystal fiber, the specific sample, the perturbed sample and the local interpretability algorithm, determine the contribution degree of the local interpretability algorithm to the first - dimensional structural parameter; if the contribution degree of the local interpretability algorithm to the first - dimensional structural parameter is basically consistent with the contribution degree of the global interpretability algorithm to the first - dimensional structural parameter, then the minimum - dimensional fiber bending resistance performance parameter corresponding to the intelligent reverse design optimization model of anti - bending photonic crystal fiber is optimal; and / or, The method for determining the contribution degree of the local interpretability algorithm to the first - dimensional structural parameter based on the intelligent reverse design optimization model of anti - bending photonic crystal fiber, the specific sample, the perturbed sample and the local interpretability algorithm includes: respectively based on the intelligent reverse design optimization model of anti - bending photonic crystal fiber, conduct first - dimensional structural parameter prediction on the specific sample and the perturbed sample to obtain a first predicted structural parameter and a second predicted structural parameter; Calculate the similarity between the first predicted structural parameter and the second predicted structural parameter, and configure the similarity as the weight of the perturbed sample; based on the perturbed sample configured with the weight, train the local interpretability algorithm to obtain the contribution degree of each - dimensional fiber bending resistance performance parameter corresponding to the minimum - dimensional fiber bending resistance performance parameter to the first - dimensional structural parameter.

8. An intelligent design system for a bend-resistant photonic crystal fiber, applied to reverse prediction tasks, is characterized in that Including: A first generation unit for generating the first - dimensional structural parameters of the anti - bending photonic crystal fiber and their corresponding second - dimensional fiber bending resistance performance parameters greater than the first - dimensional structural parameters; A first construction unit for constructing an intelligent reverse design model of anti - bending photonic crystal fiber for predicting the first - dimensional structural parameters based on the first - dimensional structural parameters and their corresponding second - dimensional fiber bending resistance performance parameters; A first calculation unit for calculating the structural parameter fitting coefficient corresponding to the intelligent reverse design model of anti - bending photonic crystal fiber; A second construction unit, configured to construct an anti-bending photonic crystal fiber intelligent inverse design optimization model not lower than the corresponding anti-bending performance parameter of the structure parameter fitting coefficient and the minimum-dimensional fiber anti-bending performance parameter among the corresponding second-dimensional fiber anti-bending performance parameters; or, Comprising: an electronic device, the electronic device is configured with a processor and a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the intelligent design method according to any one of claims 1 to 5; or, Comprising: a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the intelligent design method according to any one of claims 1 to 5 is implemented; or, Comprising: a computer program product, the computer program product is provided with a computer program / instructions, and when the computer program / instructions are executed by a processor, the intelligent design method according to any one of claims 1 to 5 is implemented.

9. Intelligent design of a bend-resistant photonic crystal fiber, applied to forward prediction tasks, characterized in that, Comprising: An intelligent design system of an anti-bending photonic crystal fiber applied to an inverse prediction task as claimed in claim 8; and, A third construction unit, configured to construct a plurality of anti-bending photonic crystal fiber intelligent forward design models for predicting the second-dimensional fiber anti-bending performance parameters by using a preset machine learning algorithm or a preset network corresponding to the intelligent design of the anti-bending photonic crystal fiber based on the first-dimensional structure parameters and the corresponding second-dimensional fiber anti-bending performance parameters; A second calculation unit, configured to calculate respectively a plurality of fiber anti-bending performance fitting coefficients corresponding to the plurality of anti-bending photonic crystal fiber intelligent forward design models; A determination unit, configured to determine an anti-bending photonic crystal fiber intelligent optimal forward design model based on the plurality of fiber anti-bending performance fitting coefficients; or, Comprising: an electronic device, the electronic device is configured with a processor and a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the intelligent design method according to claim 6; or, Comprising: a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the intelligent design method according to claim 6 is implemented; or, Comprising: a computer program product, the computer program product is provided with a computer program / instructions, and when the computer program / instructions are executed by a processor, the intelligent design method according to claim 6 is implemented.

10. An intelligent analysis system for a bending-resistant photonic crystal fiber, characterized in that, Comprising: An intelligent design system of an anti-bending photonic crystal fiber applied to an inverse prediction task as claimed in claim 8; and, A locality interpretability unit, configured to use a locality interpretability algorithm to configure each-dimensional fiber anti-bending performance parameter corresponding to the minimum-dimensional fiber anti-bending performance parameter in the test set for training the anti-bending photonic crystal fiber intelligent inverse design optimization model as a specific sample; A perturbation sample generation unit, configured to use an explanation instance function to adjust the eigenvalue corresponding to the specific sample to generate a corresponding perturbation sample; A contribution degree determination unit, configured to determine the contribution degree to the first-dimensional structure parameter determined by the locality interpretability algorithm based on the intelligent inverse design optimization model of the bending-resistant photonic crystal fiber, the specific sample, the perturbation sample, and the locality interpretability algorithm; A judgment unit, configured to determine that the minimum-dimensional fiber bending-resistant performance parameter corresponding to the intelligent inverse design optimization model of the bending-resistant photonic crystal fiber is optimal if the contribution degree to the first-dimensional structure parameter determined by the locality interpretability algorithm is basically consistent with the contribution degree to the first-dimensional structure parameter determined by the global interpretability algorithm; Or, It includes: an electronic device, the electronic device is configured with a processor and a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the intelligent analysis method according to claim 7; or, It includes: a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the intelligent analysis method according to claim 7 is implemented; or, It includes: a computer program product, the computer program product is provided with a computer program / instructions, and when the computer program / instructions are executed by a processor, the intelligent analysis method according to claim 7 is implemented.

Images