Intelligent optical computing reverse design model, method and architecture

Abstract

The invention relates to the technical field of optical computing, in particular to an intelligent optical computing reverse design model, method and architecture. The model comprises an input division module used for dividing an input electromagnetic field into an input electromagnetic field segment set based on the Huygens-Fresnel law; the wavefront prediction module is used for inputting each input electromagnetic field segment in the input electromagnetic field segment set and a structure parameter corresponding to the intelligent optical computing equipment to be designed into an electromagnetic neural local solver corresponding to the input electromagnetic field segment, the electromagnetic nerve local solver predicts an output electromagnetic field segment according to the input electromagnetic field segment and the structure parameters, and an output electromagnetic field segment set is obtained; and the output splicing module is used for splicing all the output electromagnetic field segments in the output electromagnetic field segment set based on the Huygens-Fresnel law to obtain an output electromagnetic field under global response so as to adjust the structural parameters according to the output electromagnetic field. By adopting the scheme, the reverse design of the intelligent optical computing equipment can be realized.

Description

Technical Field

[0001] The present disclosure relates to the field of optical computing technology, and in particular to an intelligent optical computing reverse design model, method and architecture. Background Art

[0002] With the rapid development of artificial intelligence and scientific computing, the complexity and scale of computing needs are also increasing. However, the existing electronic computing technology is limited by Moore's Law, and its performance is gradually approaching saturation, making it difficult to effectively cope with the increasingly stringent requirements of large-scale complex algorithms on computing power and power consumption. Light has natural advantages such as high throughput and low latency in the propagation process. Optical computing technology that uses photons instead of electrons as computing carriers is seen as the key to breaking the existing computing bottleneck. Summary of the invention

[0003] The present disclosure aims to solve one of the technical problems in the related art at least to some extent.

[0004] To this end, a first objective of the present disclosure is to propose an intelligent optical computing reverse design model to perform reverse design on intelligent optical computing devices and achieve a balance between design efficiency and accuracy.

[0005] The second objective of the present disclosure is to propose an intelligent optical computing reverse design method.

[0006] The third objective of the present disclosure is to propose an intelligent optical computing reverse design architecture.

[0007] To achieve the above objectives, the first embodiment of the present disclosure proposes an intelligent optical computing reverse design model, including: An input partitioning module, for partitioning the input electromagnetic field into a set of input electromagnetic field segments based on the Huygens-Fresnel law and according to a total wavefront width of the input electromagnetic field and a wavefront width corresponding to each electromagnetic neural local solver in the set of electromagnetic neural local solvers; A wavefront prediction module, used for inputting each input electromagnetic field segment in the input electromagnetic field segment set and the structural parameters corresponding to the intelligent optical computing device to be designed into the electromagnetic neural local solver corresponding to the input electromagnetic field segment, so that the electromagnetic neural local solver predicts the output electromagnetic field segment according to the input electromagnetic field segment and the structural parameters to obtain the output electromagnetic field segment set; The output splicing module is used to splice all the output electromagnetic field segments in the output electromagnetic field segment set based on the Huygens-Fresnel law to obtain an output electromagnetic field under a global response, so as to adjust the structural parameters according to the output electromagnetic field.

[0008] Optionally, the input division module is used to divide the input electromagnetic field into a set of input electromagnetic field segments according to the total wavefront width of the input electromagnetic field and the wavefront width corresponding to each electromagnetic neural local solver in the set of electromagnetic neural local solvers, specifically for: The total wavefront width of the input electromagnetic field is divided based on the wavefront width corresponding to each electromagnetic neural local solver in the electromagnetic neural local solver set to obtain a division result; If the division result indicates that there is an input electromagnetic field segment with a width smaller than the wavefront width in the input electromagnetic field, zero padding is added to the end side and / or the end side of the input electromagnetic field until the width of each input electromagnetic field segment is the wavefront width, thereby obtaining an input electromagnetic field segment set, wherein the input electromagnetic field segment is regarded as a secondary wave generated by a central source point on the input electromagnetic field segment.

[0009] Optionally, when the output splicing module is used to splice all the output electromagnetic field segments in the output electromagnetic field segment set, it is specifically used to: All output electromagnetic field segments in the set of output electromagnetic field segments are coherently superimposed, wherein the output electromagnetic field segments are regarded as envelope surfaces generated by secondary wave sources on the output electromagnetic field segments.

[0010] Optionally, when the electromagnetic neural local solver is used to predict the output electromagnetic field segment according to the input electromagnetic field segment and the structural parameter, it is specifically used to: Performing complex linear connection on the input electromagnetic field segments to represent the non-physical information propagation process of the input electromagnetic field segments in the intelligent optical computing device to be designed, and obtaining the input electromagnetic field segments after non-physical propagation; Performing real number nonlinear connection on the structural parameters, and aligning information between the connected structural parameters and the input electromagnetic field segment after non-physical propagation, so as to represent the physical information propagation process of the input electromagnetic field segment after propagation in the intelligent optical computing device to be designed, and obtaining the input electromagnetic field segment after physical propagation; The input electromagnetic field segments after physical propagation are connected in a complex linear manner to represent the non-physical information propagation process of the input electromagnetic field segments after physical propagation in the intelligent optical computing device to be designed, and output electromagnetic field segments are obtained.

[0011] Optionally, when the electromagnetic neural local solver is used for performing complex linear connection, it is specifically used for: Acquire a first weight matrix and a first intercept term used in performing complex linear connection, and perform complex linear connection based on the first weight matrix and the first intercept term, wherein the first weight matrix and the first intercept term are obtained by training an initial electromagnetic neural local solver.

[0012] Optionally, when the electromagnetic neural local solver is used for performing real number nonlinear connection, it is specifically used for: A second weight matrix, a second intercept term and a nonlinear operation function used in performing real nonlinear connection are obtained, and a complex linear connection is performed based on the second weight matrix, the second intercept term and the nonlinear operation function, wherein the second weight matrix, the second intercept term and the nonlinear operation function are obtained by training an initial electromagnetic neural local solver.

[0013] Optionally, before inputting each input electromagnetic field segment in the set of input electromagnetic field segments and the structural parameters corresponding to the intelligent optical computing device to be designed into the electromagnetic neural local solver corresponding to the input electromagnetic field segment, the wavefront prediction module is further used to: Acquire a training sample set, wherein each training sample in the training sample set includes an input electromagnetic field training sample and a structural parameter training sample; Performing gradient back propagation training on the initial electromagnetic neural local solver based on the training sample set until an electromagnetic neural local solver that meets the training requirements is obtained; Acquire a test sample set, wherein each of the test samples includes an input electromagnetic field test sample and a structural parameter test sample; The electromagnetic neural local solver is tested based on the test sample set. If the test result does not meet the test requirements, the electromagnetic neural local solver is re-trained by gradient back propagation until an electromagnetic neural local solver whose test result meets the test requirements is obtained.

[0014] Optionally, when the wavefront prediction module is used to perform gradient back propagation training on the initial electromagnetic neural local solver based on the training sample set, it is specifically used to: In the process of performing gradient back-propagation training on the initial electromagnetic neural local solver based on the training sample set, any training sample in the training sample set is input into the initial electromagnetic neural local solver to obtain a predicted value corresponding to any training sample, and a first error value between the predicted value and a true value corresponding to any training sample is determined, and the network parameters of the initial electromagnetic neural local solver are updated based on the first error value to minimize the loss function.

[0015] Optionally, before the wavefront prediction module is used to determine a first error value between the predicted value and a true value corresponding to any one of the training samples, it is further used to: Input any of the training samples into a numerical calculation solver to obtain a true value corresponding to any of the training samples.

[0016] To achieve the above-mentioned purpose, a second aspect of the present disclosure provides an intelligent optical computing reverse design method, including: Acquire a sample data set, wherein the sample data set includes an input electromagnetic field sample; Based on the sample data set and the intelligent optical computing reverse design model shown in any one of the first aspects, training the structural parameters of the intelligent optical computing device to be designed to obtain target structural parameters that meet the training requirements; An intelligent optical computing device corresponding to the target structural parameters is constructed.

[0017] Optionally, the training of the structural parameters of the intelligent optical computing device to be designed to obtain target structural parameters that meet training requirements includes: Inputting the structural parameters of the intelligent optical computing device to be designed and any input electromagnetic field sample in the sample data set into the intelligent optical computing reverse design model to obtain a predicted output electromagnetic field; Inputting the structural parameters and any input electromagnetic field sample into a numerical calculation solver to obtain a real output electromagnetic field corresponding to any input electromagnetic field sample; Determine a second error value between the predicted output electromagnetic field and the actual output electromagnetic field. If the second error value is greater than an error threshold, update the structural parameters until the second error value is no greater than the error threshold, thereby obtaining target structural parameters that meet training requirements.

[0018] To achieve the above objectives, the third aspect of the present disclosure proposes an intelligent optical computing reverse design architecture, including: A sample acquisition unit, used to acquire a sample data set, wherein the sample data set includes an input electromagnetic field sample; A parameter training unit, configured to train the structural parameters of the intelligent optical computing device to be designed based on the sample data set and the intelligent optical computing reverse design model shown in any one of the first aspects, to obtain target structural parameters that meet the training requirements; The device construction unit is used to construct an intelligent optical computing device corresponding to the target structural parameters.

[0019] In summary, the intelligent optical computing reverse design model, method and architecture provided by the present invention, by adopting an electromagnetic neural local solver as a local electromagnetic field solver and using the Huygens-Fresnel global method to connect the predicted local output electromagnetic field, can directly, quickly and accurately predict the full-wave electromagnetic field distribution based on arbitrarily changing input fields and sub-wavelength structures of arbitrary sizes, thereby realizing the reverse design of the structural parameters of the intelligent optical computing device to be designed based on the predicted output electromagnetic field.

[0020] Additional aspects and advantages of the present disclosure will be given in part in the following description and in part will be obvious from the following description or learned through practice of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and / or additional aspects and advantages of the present disclosure will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which: Figure 1 A schematic diagram of the structure of an intelligent optical computing reverse design model provided by an embodiment of the present disclosure; Figure 2 A design concept diagram of an intelligent optical computing reverse design model provided by an embodiment of the present disclosure; Figure 3 A conceptual diagram of a local solver for an electromagnetic field fast simulation operator provided by an embodiment of the present disclosure; Figure 4 A schematic diagram of a flow chart of an intelligent optical computing reverse design method provided by an embodiment of the present disclosure; Figure 5 A schematic diagram of the structure of an intelligent optical computing reverse design architecture provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0022] Embodiments of the present disclosure are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present disclosure, and should not be construed as limiting the present disclosure.

[0023] Optical computing hardware is expected to drive the development of artificial intelligence in the future due to its advantages over electronic implementations in terms of low latency, minimal power consumption, high parallelism, and high speed, as well as the fact that it is a form of analog computing that eliminates the need for digital-to-analog conversion. Currently, optical neural networks (optical intelligent computing in the analog domain) are mainly implemented through interference based on Mach-Zehnder interferometers (MZIs), diffraction using trained modulation units, and wavelength division multiplexing implemented by microrings. However, due to their large footprint, the integration density and scale of these optical neural networks are significantly limited.

[0024] Artificially designed subwavelength structures, known as subwavelength structures, may significantly improve the level of integration due to their subwavelength, super-atomic size. The interaction between light and subwavelength structures enables phase and amplitude modulation of the light wavefront, which has wide applications in focusing, imaging, vortex generation, spectroscopy, and other fields. Compared with traditional spatial light diffraction plates, subwavelength structures have more compact sizes, high integration capabilities, and can realize on-chip all-optical diffraction neural networks by cascading multiple levels of subwavelength structures. The diffraction optical computing supersystem utilizes multi-layer subwavelength structure structures to flexibly modulate the wavefront, and is expected to become the next-generation computing hardware platform with its compact size, high computing density, and low power consumption.

[0025] Early subwavelength structure design mainly relied on intuitive manual parameter adjustment or prior knowledge through the use of analytical models, which resulted in a large number of trial-and-error processes, making the design process inefficient and arduous. In recent years, inverse design has provided a new paradigm for device design, which optimizes the design goals by adjusting the structural parameters to achieve the desired optical response, greatly improving the design efficiency and device performance. The optimization algorithms mainly include heuristic algorithms, gradient optimization, etc. These methods usually use iterative numerical simulation algorithms such as finite element method (FEM), finite difference time domain (finite difference time domain, numerical calculation method) and rigorous coupled wave analysis (rigorous coupled-wave analysis, RCWA) to obtain forward evaluation of the device. These methods solve electromagnetic field constraints and Maxwell equations by replacing differentials with differences, providing accurate and reliable results, but with high computational costs in terms of memory usage and computational time. The inefficiency of numerical methods may cause the inverse design process to be time-consuming and costly because the forward evaluation is repeated in step-by-step iterations. In order to reduce the computational burden, neural networks are used as an alternative to numerical simulation methods to accelerate inverse design. In recent years, various deep learning models have been combined with inverse design to achieve better performance and higher efficiency, including discriminative architectures such as fuzzy clustering neural network (FCNN), convolutional neural network (CNN) and recurrent neural network (RNN), as well as generative models such as variational autoencoder (VAE) and generative adversarial network (GAN). These models have complex architectures and a large number of connections and weights, showing strong learning capabilities, but at the same time consuming huge computing resources. In essence, these methods are not specifically designed to deal with the interaction of electromagnetic waves with subwavelength structures, resulting in model redundancy and waste of computing resources. In addition, there are other specific limitations of previous designs using deep learning. These works only focus on indirect properties such as spectrum or transmittance, rather than the light field itself. In addition, these methods lack flexibility because they ignore variable input fields and limit the device size to a fixed value. In addition, in large-scale design, the design space has extremely high degrees of freedom (DOF), which requires a very large number of simulations and optimizations. As the device size increases and the functions become more complex, all these issues pose considerable challenges to reverse design.

[0026] For example, when light propagates in the transverse electric wave (TE) mode in the basic slab mode, the on-chip propagation process is modeled by two parts: the slab mode propagation model and the optical phase modulation model. Among them, the slab mode propagation model has been discussed in detail and well verified in many previous studies, and its calculation results are almost completely consistent with the results of the numerical calculation method solver. The slab mode propagation adopted is derived based on the Fourier optics method and can be expressed as:

[0027] in, represents the optically phase modulated electric field, and are discrete Fourier transform and inverse discrete Fourier transform, respectively, and is the filter matrix: In actual computer calculations, the electric field is a string of discrete values ​​uniformly distributed in space z, which can be expressed as:

[0028] Discrete Fourier transform of these data can obtain the spatial frequency component of the magnetic field along the z direction:

[0029] Filter Matrix The value of can be obtained by numerical simulation, which can be obtained by looking up the table:

[0030] The final calculation result of the optical phase modulation electric field is:

[0031] Previously, physical modeling errors of on-chip optical networks were mainly attributed to the optical phase modulation model. Currently, optical phase modulation can be modeled by analytical models or numerical simulations. Analytical models are derived based on simplifications of physical processes, such as the effective refractive index method and the Fourier optics method. Although these methods can quickly calculate the output field distribution, they cannot guarantee the accuracy of the modeling. Although numerical simulations can ensure accurate modeling of designed devices, the device optimization time and computational cost will be extremely high for large-scale, multi-layer cascaded and high-degree-of-freedom computing supersystems.

[0032] In addition, unlike the design of metalenses or some single-layer subwavelength structures, the cascading of multi-layer structures leads to the accumulation of physical modeling errors, which has a significant impact on the effectiveness of analog computing systems. The accuracy of existing advanced analytical models cannot meet the requirements of large-scale computing supersystems. Therefore, the performance of devices designed using analytical methods is significantly degraded when deployed in actual physical experiments, which requires additional complex calibration to improve performance. The lack of reliable and efficient physical modeling tools has hindered the development of large-scale optical computing in performing complex tasks. There are currently no tools or models that can reliably and efficiently deal with such problems.

[0033] That is, on-chip computing supersystems composed of multi-layer subwavelength structures are expected to become the next generation of computing hardware, with light-speed processing capabilities and low power consumption, but the current design paradigm has hindered its development. So far, neither numerical methods nor analytical analysis methods have been able to strike a balance between efficiency and accuracy in design.

[0034] The present disclosure is described in detail below with reference to specific embodiments.

[0035] Figure 1 The structure diagram of an intelligent optical computing reverse design model provided by the embodiment of the present disclosure is shown in FIG. Figure 1 As shown, the intelligent optical computing reverse design model includes: An input partitioning module, for partitioning the input electromagnetic field into a set of input electromagnetic field segments based on the Huygens-Fresnel law and according to a total wavefront width of the input electromagnetic field and a wavefront width corresponding to each electromagnetic neural local solver in the set of electromagnetic neural local solvers; A wavefront prediction module is used to input each input electromagnetic field segment in the input electromagnetic field segment set and the structural parameters corresponding to the intelligent optical computing device to be designed into the electromagnetic neural local solver corresponding to the input electromagnetic field segment, so that the electromagnetic neural local solver predicts the output electromagnetic field segment according to the input electromagnetic field segment and the structural parameters to obtain the output electromagnetic field segment set; The output splicing module is used to splice all output electromagnetic field segments in the output electromagnetic field segment set based on the Huygens-Fresnel law to obtain the output electromagnetic field under the global response, so as to adjust the structural parameters according to the output electromagnetic field.

[0036] According to some embodiments, the electromagnetic neural local solver refers to a local solver in an electromagnetic neural network (Electromagnetic neural networks, electromagnetic field fast simulation operator), that is, an electromagnetic field fast simulation operator local solver. The electromagnetic field fast simulation operator local solver is a neural network with a physics-inspired architecture that can predict the output electromagnetic field based on a given input electromagnetic field and structural parameters. The electromagnetic field fast simulation operator local solver is a key building block for the intelligent optical computing reverse design model to effectively predict the input electromagnetic field. It considers the variable input field in the reverse design network for the first time. Through physically inspired modifications, it can successfully perform full-wave prediction accurately and efficiently, and can accurately predict the electromagnetic field distribution based on arbitrarily changing input fields, and exhibits good interpretability and generalization capabilities.

[0037] In some embodiments, the structural parameters refer to the structural parameters of the sub-wavelength structure used by the intelligent optical computing device to be reverse-designed when modulating the input light. The intelligent optical computing device may be a device corresponding to a computing super system. The computing super system includes a linear system and a nonlinear system.

[0038] In some embodiments, the wavefront width corresponding to the local solver of the electromagnetic field fast simulation operator can be adjusted according to the actual application scenario. For example, the wavefront width corresponding to the local solver of the electromagnetic field fast simulation operator can be 0.6 micrometers.

[0039] It should be noted that when the total wavefront width of the input electromagnetic field is significantly larger than the wavefront width corresponding to the local solver of the electromagnetic field fast simulation operator, for example, in the upcoming follow-up task, the total wavefront width of the designed input electromagnetic field is 100 microns, and the wavefront width that the local solver of the electromagnetic field fast simulation operator can handle is 0.6 microns. The total wavefront width of the input electromagnetic field far exceeds the wavefront width that a single electromagnetic field fast simulation operator local solver can handle, reaching a difference of three orders of magnitude. If a single neural network is directly applied to a 100-micron-long device, many problems will arise. First, the complexity of the problem rises sharply. As the application area grows linearly, the cost of training the network, including generating data sets, increases multiplicatively. In other words, the computational cost required to train a single forward prediction network for a 100-micron-wide device is at least several orders of magnitude greater than that of training a network for a 0.6-micron-wide device. In addition, even if a lot of effort is put into training this single neural network, the network lacks flexibility and versatility, that is, once the device size changes, a new network will have to be trained from scratch.

[0040] Therefore, inspired by the principle of optical diffraction propagation, a global method with physical information is proposed to seamlessly connect multiple local solvers of electromagnetic field fast simulation operators. The global method with physical information is a strategy to extend the forward model to arbitrarily large sizes. It is a physically inspired strategy. Inspired by the Huygens-Fresnel law, it can successfully physically splice the local responses predicted by the local solvers of electromagnetic field fast simulation operators together to form a global response to obtain arbitrarily large electromagnetic field distributions. This physically inspired strategy enables the model to be flexibly applied to devices of non-fixed size, so that the model can handle inverse design problems of any scale without increasing the complexity of the problem or reducing the design performance. More importantly, this method aligns the process of wavefront splicing with the real physical process, significantly reducing the splicing error compared to previous non-physical inspired methods. In addition, in previous methods, the network structure lacks scalability, which limits the ability of the algorithm to adjust parameters in a fixed-size region. However, with the introduction of the global method with physical information, the network shows strong scalability. Taking the inverse design of optical neural networks as an example, the intelligent optical computing inverse design model can design optical intelligent computing in simulation domains of arbitrary width and depth.

[0041] It should be noted that the intelligent optical computing reverse design model provided by the embodiment of the present disclosure adopts the electromagnetic field fast simulation operator local solver as the local electromagnetic field solver, and uses the global method containing physical information to connect the predicted local output electromagnetic field. It can directly, quickly and accurately predict the full-wave electromagnetic field distribution based on arbitrarily changing input fields and sub-wavelength structures of arbitrary sizes, so that the reverse design of the structural parameters of the intelligent optical computing device to be designed can be realized based on the predicted output electromagnetic field. More importantly, the previous deep learning framework predicts indirect physical quantities, such as transmittance or spectrum, to meet the requirements of specific tasks. In contrast, the intelligent optical computing reverse design model can directly predict the electromagnetic field components. If transmittance is required, it can be calculated through the electromagnetic field components. The scalability and flexibility of the intelligent optical computing reverse design model meet the unprecedented reverse design needs and can cope with any scale, high degree of freedom and complex functions, which exceeds the capabilities of other deep learning-based methods.

[0042] Optionally, Figure 2 This is a design concept diagram of an intelligent optical computing reverse design model provided by an embodiment of the present disclosure. Figure 2 As shown in , the device is optimized through iterative reasoning and tuning to produce the desired electromagnetic response. The large scale, high design freedom and complex functionality of the device, as well as the challenging forward modeling, are the main obstacles faced by the reverse design. The proposed model aims to address these issues.

[0043] It should be noted that in the global method containing physical information, according to the Huygens-Fresnel law, during the propagation of the light field, each point on the wavefront can be regarded as a central source, generating secondary waves; then, these secondary waves are coherently superimposed to form a new wavefront. In other words, Figure 2 As shown, The input division module can divide the total wavefront width of the input electromagnetic field based on the wavefront width corresponding to each electromagnetic neural local solver in the electromagnetic neural local solver set to obtain a division result; if the division result indicates that there is an input electromagnetic field segment with a width less than the wavefront width in the input electromagnetic field, zero padding is added to the end side and / or the end side of the input electromagnetic field until the width of each input electromagnetic field segment is the wavefront width, thereby obtaining an input electromagnetic field segment set, wherein the input electromagnetic field segment is regarded as a secondary wave generated by a central source point on the input electromagnetic field segment.

[0044] The output splicing module can coherently superimpose all output electromagnetic field segments in the output electromagnetic field segment set, wherein the output electromagnetic field segment is regarded as an envelope surface generated by the secondary wave source on the output electromagnetic field segment.

[0045] For example, first, a wavefront with a width of 100 microns can be divided into multiple segments with a width of 0.6 microns. Each segment can be regarded as an independent point on the wavefront to generate a secondary wave. Then, each 0.6-micron wavefront segment and its corresponding structural parameters with a width of 3.0 microns are input into the local solver of the electromagnetic field fast simulation operator, which will generate output electromagnetic field segments with a width of 3.0 microns. These output electromagnetic field segments represent the envelope surface generated by the secondary wave source. Finally, the predicted output electromagnetic field segments are coherently superimposed to obtain the final wavefront modulation result of the 100-micron wide diffraction subwavelength structure on the input field.

[0046] According to some embodiments, Figure 3 This is a conceptual diagram of the design of a local solver for electromagnetic field fast simulation operators provided by the embodiment of the present disclosure. Figure 3 As shown in , it shows the interpretable correspondence between the local solver architecture of the electromagnetic field fast simulation operator and the light propagation process. The physical process of wavefront modulation can be divided into three stages, of which the first and third stages are non-physical information propagation stages, and the second stage is the physical information propagation stage. When the electromagnetic neural local solver is used to predict the output electromagnetic field segment based on the input electromagnetic field segment and structural parameters, it is specifically used for: Performing complex linear connection on the input electromagnetic field segments to represent the non-physical information propagation process of the input electromagnetic field segments in the intelligent optical computing device to be designed, and obtaining the input electromagnetic field segments after non-physical propagation; Performing real number nonlinear connection on the structural parameters, and aligning the connected structural parameters with the input electromagnetic field segment after non-physical propagation, so as to represent the physical information propagation process of the input electromagnetic field segment after propagation in the intelligent optical computing device to be designed, and obtaining the input electromagnetic field segment after physical propagation; The input electromagnetic field segments after physical propagation are connected in a complex linear manner to represent the non-physical information propagation process of the input electromagnetic field segments after physical propagation in the intelligent optical computing device to be designed, and the output electromagnetic field segments are obtained.

[0047] According to some embodiments, the information alignment method includes but is not limited to other mathematical operations such as element-by-element multiplication, concatenation, outer product, and attention mechanism. Among them, the element-by-element multiplication operation corresponds to the physical process of optical phase modulation, and can be replaced by other mathematical operations according to the specific physical process required to be designed in the intelligent optical computing device to be designed; concatenation is to connect two vectors along a certain dimension to retain the original electromagnetic field information; outer product refers to generating the outer product matrix of two vectors to capture the second-order interaction between features; attention mechanism uses the Transformer model to align different modal information. In this embodiment, the attention mechanism can be used to align electromagnetic field information with sub-wavelength physical structure information.

[0048] For example, you can use [ x , y ] represents information alignment operation on two complex signals. In this embodiment, x and y They are the structural parameters after connection and the input electromagnetic field segments after non-physical propagation, respectively.

[0049] In some embodiments, the non-physical information propagation process refers to, for example, a planar mode propagation process of an input signal outside a sub-wavelength structure in a to-be-designed intelligent optical computing device (although the distance is short, it does exist).

[0050] In some embodiments, the physical information propagation process refers to, for example, an optical phase modulation process in which an input signal in a to-be-designed intelligent optical computing device passes through a sub-wavelength structure.

[0051] According to some embodiments, in electromagnetic field signal processing (such as communication, radar, electromagnetic compatibility analysis, etc.), electromagnetic field data is usually represented in complex form (including amplitude and phase information). Complex linear connection is a key operation that maps complex features to a unified semantic space. Its core is to realize information processing through linear transformation in the complex domain.

[0052] In some embodiments, when the electromagnetic neural local solver is used to perform complex linear connection, it can obtain the first weight matrix and the first intercept term used in the complex linear connection, and perform the complex linear connection based on the first weight matrix and the first intercept term.

[0053] For example, if z 1 is the output complex signal, is the first weight matrix, x 1 is the input signal for complex linear connection, 1 is the first intercept term, then the complex linear connection can be generally expressed as:

[0054] According to some embodiments, when real number nonlinear connections are performed on structural parameters, complex interactions between parameters can be modeled and high-dimensional complex information can be captured through nonlinear transformation.

[0055] In some embodiments, when the electromagnetic neural local solver is used to perform real nonlinear connection, it can obtain the second weight matrix, second intercept term and nonlinear operation function used when performing real nonlinear connection, and perform complex linear connection based on the second weight matrix, second intercept term and nonlinear operation function.

[0056] For example, if 2 is the output nonlinear signal, 2 is the input signal for real nonlinear connection, is the second weight matrix, 2 is the second intercept term, is a nonlinear operation function, then the real nonlinear connection can be generally expressed as:

[0057] It should be noted that, unlike the previous practice of ignoring the input waveform, the local solver of the electromagnetic field fast simulation operator takes into account the changes in the input electromagnetic field. Based on the dual-input architecture design, the local solver of the electromagnetic field fast simulation operator can simultaneously predict the output electromagnetic field results based on the input electromagnetic field and structural parameters. In addition, it is worth noting that, in general, neural networks are regarded as black boxes that lack interpretability. However, in the local solver of the electromagnetic field fast simulation operator, the input electromagnetic field and optical phase modulation are encoded in complex form instead of processing the real and imaginary parts separately. Different parts of the network are connected in different ways. The input electromagnetic field and the output electromagnetic field are linearly connected through complex weights, while the input structural parameters are nonlinearly connected through real weights. This physically inspired neural network is carefully designed to be physically interpretable and has a convincing explanation of the physical meaning contained in the network architecture, which can show strong physical interpretability.

[0058] According to some embodiments, the electromagnetic field fast simulation operator local solver can be obtained by pre-training the initial electromagnetic field fast simulation operator local solver. That is, the wavefront prediction module can also be used for: Acquire a training sample set, wherein each training sample in the training sample set includes an input electromagnetic field training sample and a structural parameter training sample; Performing gradient back propagation training on the initial electromagnetic neural local solver based on the training sample set until an electromagnetic neural local solver that meets the training requirements is obtained; Acquire a test sample set, wherein each test sample in the test samples includes an input electromagnetic field test sample and a structural parameter test sample; The electromagnetic neural local solver is tested based on the test sample set. If the test result does not meet the test requirements, the electromagnetic neural local solver is re-trained by gradient back propagation until an electromagnetic neural local solver whose test result meets the test requirements is obtained.

[0059] In some embodiments, Figure 3 As shown in , the wavefront prediction module is used to perform gradient back propagation training on the initial electromagnetic neural local solver based on the training sample set, specifically for: In the process of gradient back propagation training of the initial electromagnetic neural local solver based on the training sample set, any training sample in the training sample set is input into the initial electromagnetic neural local solver to obtain a predicted value corresponding to any training sample, and a first error value between the predicted value and the true value corresponding to any training sample is determined, and the network parameters of the initial electromagnetic neural local solver are updated based on the first error value to minimize the loss function.

[0060] The true value corresponding to any training sample can be obtained by inputting any training sample into a numerical calculation solver. The numerical calculation solver can be, for example, a finite difference time domain method solver.

[0061] It should be noted that when the initial electromagnetic neural local solver is trained with gradient back propagation based on the training sample set, the training parameters include but are not limited to the parameters used in the above-mentioned complex linear connection and real nonlinear connection. In other words, , , , 1, 2 can be obtained by training the initial electromagnetic neural local solver.

[0062] Taking a scenario as an example, when a single electromagnetic field fast simulation operator local solver is capable of handling the wavefront modulation problem in a limited area with a width of 900 nanometers, the input-output pairs can be simulated and generated using numerical calculation methods after random parameterization. A total of 90,000 input-output pairs are used to constitute the training sample set and test sample set of the network, of which 80,000 are used as training sample sets and 10,000 are used as test sample sets. The training learning curve shown indicates that the test error converges after more than 70 training cycles.

[0063] It is worth noting that although the simulation of the above network dataset is quite large and requires a lot of computing resources, compared with those methods that require repeated access to computationally expensive and time-consuming numerical calculation method solvers, this simulation work can be regarded as a one-time investment. In other words, the electromagnetic field fast simulation operator local solver can serve as a fast-response alternative to inefficient numerical simulations.

[0064] According to some embodiments, the intelligent optical computing device to be reverse-engineered includes a single-layer metal wire composed of meta-atoms, and the structural parameters corresponding to these single-layer metal wires are the structural parameters corresponding to the intelligent optical computing device, wherein the thickness of the silicon film is 220 nanometers, the width of the meta-atom is 100 nanometers, and the period of the diffraction metal wire is 200 nanometers. The amplitude and phase modulation of the light field by the intelligent optical computing device can be controlled by changing the gap length.

[0065] In some embodiments, thanks to the modification of physics-inspired methods, the local solver of the electromagnetic field fast simulation operator also exhibits excellent generalization ability. Under the conditions of the same network depth, comparable number of parameters and the same training sample set, the local solver of the electromagnetic field fast simulation operator has stronger learning ability than the non-physics-inspired network. Figure 3 As shown in , it shows the wavefront profile of the output electromagnetic field in the yz plane calculated by the local solver of the electromagnetic field fast simulation operator, the numerical calculation method solver and the traditional non-physical inspiration method on the same test sample. The test sample is generated differently from any training sample in the training sample set. Among them, the local solver of the electromagnetic field fast simulation operator successfully predicts the results consistent with the simulation results of the numerical calculation method, while the non-physical inspiration method fails to achieve the same accuracy. The local solver of the electromagnetic field fast simulation operator driven by physical information performs better in the prediction accuracy of both amplitude and phase, because the local solver of the electromagnetic field fast simulation operator essentially satisfies the linear response constraints of the physical system.

[0066] In summary, the intelligent optical computing reverse design model provided by the embodiment of the present disclosure integrates deep learning technology with basic physical principles and integrates physical heuristic correction into deep learning to obtain powerful physical interpretability and generalization ability, achieve better performance, and can show excellent interpretability and generalization ability beyond traditional networks, and can realize reliable, efficient and flexible reverse design of intelligent optical computing devices with large-scale, high degree of freedom and complex functions, and achieve a balance between the accuracy and efficiency of the design. In addition, considering that the electromagnetic field itself is a composite field with phase and amplitude, the electromagnetic field information and optical phase modulation information are encoded as complex tensors; in addition, considering that light propagation can be regarded as a linear process of Fourier filtering, a network structure with complex value linear connection is introduced in the local solver of the electromagnetic field fast simulation operator. These physical inspired modifications can help the local solver of the electromagnetic field fast simulation operator extract the basic physical information hidden in the data, so that the local solver of the electromagnetic field fast simulation operator can essentially meet some physical constraints of the system, which gives the local solver of the electromagnetic field fast simulation operator a generalization ability beyond the traditional network architecture and ensures high-precision prediction ability.

[0067] The disclosed embodiment also provides a method for reverse design of intelligent optical computing, which can be applied to the reverse design architecture of intelligent optical computing.

[0068] For example, Figure 4 As shown, the intelligent optical computing reverse design method includes the following steps: S101, obtaining a sample data set; According to some embodiments, a sample data set refers to a set consisting of at least one input electromagnetic field sample.

[0069] In some embodiments, the sample data set may be obtained by preprocessing original sample data in the original sample data set into wavefront data.

[0070] S102, training the structural parameters of the intelligent optical computing device to be designed based on the sample data set and the intelligent optical computing reverse design model to obtain target structural parameters that meet the training requirements; According to some embodiments, the intelligent optical computing reverse design model is the model provided by the aforementioned embodiments, and the aforementioned explanation of the intelligent optical computing reverse design model embodiment is also applicable to the intelligent optical computing reverse design model of this embodiment, which will not be repeated here.

[0071] In some embodiments, during the process of training the structural parameters of the intelligent optical computing device to be designed, the structural parameters of the intelligent optical computing device to be designed and any input electromagnetic field sample in the sample data set can be input into the intelligent optical computing reverse design model to obtain a predicted output electromagnetic field; the structural parameters and any input electromagnetic field sample are input into a numerical calculation solver to obtain a true output electromagnetic field corresponding to any input electromagnetic field sample; the second error value between the predicted output electromagnetic field and the true output electromagnetic field is determined, and if the second error value is greater than an error threshold, the structural parameters are updated until the second error value is no greater than the error threshold, thereby obtaining target structural parameters that meet the training requirements.

[0072] S103, constructing an intelligent optical computing device corresponding to the target structural parameters.

[0073] It should be noted that, at present, due to the simple architecture, analog domain optical intelligent computing can only cope with relatively basic tasks. Only by expanding the scale and increasing the complexity of the optical network can more intelligent optical computing tasks be performed. Typical methods include increasing the number of neurons, the number of layers, and introducing nonlinear activation. However, since optical intelligent computing is essentially the realization of analog computing, simulation errors in optical computing pose the biggest obstacle to expanding the scale of optical networks and realizing more complex functions. Nevertheless, by introducing the intelligent optical computing reverse design method provided by the embodiment of the present disclosure, simulation errors can be effectively solved even under large-scale and nonlinear computing architectures.

[0074] Taking a scenario as an example, the embodiment of the present disclosure provides a three-layer analog domain optical intelligent computing task for handwritten digit classification tasks. Among them, the on-chip diffraction super system is reverse-engineered to perform handwritten digit recognition tasks, including preprocessing, optical phase modulation, and output conversion; that is, before entering the optical network, the data undergoes some preprocessing steps, including downsampling the grayscale image of (28,28) to (4,4), and then flattening it into a (16,) vector; the on-chip optical neural network for optical phase modulation consists of three layers of diffractive subwavelength structures, each layer is 300 microns wide, with 1,000 independently controllable optical network neurons, and the distance between the input surface and each layer and between the layer and the output layer is fixed at 60 microns. The handwritten digital signal is encoded as the amplitude of the input light field and input into the planar waveguide through 16 reverse tapered waveguides. The light field propagates through the waveguide and then passes through the modulation of the three layers of diffractive subwavelength structures in turn. Finally, the light field reaches the output surface, where there are 10 output reverse tapered waveguides. The 10 intensity values ​​are recorded through the output ports, and the classification result is determined by the highest intensity value among all 10 output ports.

[0075] Among them, the three-layer analog domain optical intelligent computing model is a combination of the above-mentioned optical phase modulation model and the flat plate mode propagation model. The MNIST data set can be used as the original sample data set to train the three-layer analog domain optical intelligent computing, and the MNIST data set is divided into a training set containing 50,000 samples and a test set containing 10,000 samples. During the training process, the intelligent optical computing reverse design model can be used as a forward propagation model to process the input signal, and the structural parameters of the optical network are input into the intelligent optical computing reverse design model. The actual optical response of the network is propagated and calculated, and the calculated response is compared with the target optical response, and the mean square error (MSE) is used as the loss function of the optical network. In addition, the loss function can also be regarded as the design objective function of the reverse design task. The optimization process is to adjust the structural parameters of the optimized subwavelength structure by back-propagating the gradient. In this process, the three-layer analog domain optical intelligent computing can be deployed on TensorFlow, and the Adam optimizer is selected as the optimizer for this task.

[0076] In some embodiments, the intelligent optical computing reverse design model is replaced by a numerical calculation method solver and a Fourier optical method model to compare their prediction capabilities. Among them, the Fourier optical method model is considered to be the best performing analytical model to date, and the results calculated by the Fourier optical method can be regarded as a benchmark. The numerical calculation method is a very reliable full-wave numerical method for electromagnetic field simulation, and its simulation results can be regarded as true values.

[0077] According to some embodiments, the embodiments of the present disclosure provide a six-layer analog domain optical intelligent computing for voice command recognition tasks. For example, it shows a more complex on-chip optical diffraction computing architecture, which extends the analog domain optical intelligent computing to six layers and introduces three nonlinear layers in the network. The architecture of the six-layer analog domain optical intelligent computing includes preprocessing, optical phase modulation, nonlinear activation and output conversion. It shows the performance comparison results of linear analog domain optical intelligent computing and nonlinear analog domain optical intelligent computing in voice command classification. The results show that after the introduction of nonlinearity, the learning ability of the network can be enhanced, thereby improving the accuracy of classification.

[0078] According to some embodiments, for the application of optical computing in audio classification tasks, such as vowel classification, these tasks involve a small number of categories and low feature dimensions, so as to be compatible with a completely linear optical computing framework. Therefore, the embodiment of the present disclosure can select a language instruction data set as the original sample data set, which contains a total of 34 categories of speech data, of which ten categories can be selected for the classification task of six-layer analog domain optical intelligent computing. The language instruction data set has a total of 20,000 samples, of which 16,000 samples are used as training sets and 4,000 samples are used as test sets. Each sample is usually a speech segment of about 1 second in length, with a sampling rate of 16,000 Hz. These speech samples are recorded and sorted from different speakers using various recording devices, so accents, background noise, and device-specific artifacts pose great challenges to the classification task. For such a complex task, a simple shallow linear optical network is obviously not enough. Therefore, a deep nonlinear optical network is needed to solve this problem. At the same time, the intelligent optical computing inverse design method proposed in the present disclosure is used to improve the accuracy of the physical modeling of the six-layer analog domain optical intelligent computing, and enhance the efficiency and fidelity of the design.

[0079] In some embodiments, the ten selected categories. A numerical calculation method solver, a Fourier optical method model, and an intelligent optical computing reverse design model are respectively used to perform forward reasoning on the sample data with the input true label "eight". The intelligent optical computing reverse design method provided by the present disclosure achieves sufficient accuracy in physical modeling. The classification accuracy verified by the numerical calculation method solver can reach 80%, even if the depth and complexity of the network are increased. The classification accuracy of the Fourier optical method model is only 16%, which is slightly higher than random guessing.

[0080] It should be noted that when the intelligent optical computing reverse design method is applied to achieve reliable and efficient reverse design, it can provide a more accurate and faster electromagnetic field prediction tool. On the one hand, due to the simplification and abstraction of the actual physical process, the analytical model can quickly calculate the approximate results, however, its accuracy is not high, and after a multi-layer structure, the error will accumulate, resulting in a decrease in the fidelity of the final device performance; in addition, some analytical models, such as the Fourier optical method model, use a table lookup process for solution, which makes it difficult to implement gradient back propagation. On the other hand, the numerical method solves the time-dependent partial differential equations through the difference method, providing accurate results, but at the cost of significant spatial and temporal computing resources. Using data-driven deep learning methods to learn the diffraction process can eliminate the time-consuming and resource-intensive numerical simulation process, thereby achieving direct and accurate prediction of the electromagnetic field. In addition, the end-to-end learning architecture does not require specialized physical knowledge or analytical capabilities.

[0081] Based on the intelligent optical computing reverse design model, the computational super system that performs handwritten digit recognition and voice command recognition can be reverse designed reliably and efficiently. After verification by numerical calculation methods, it has been proved that it has high efficiency and high fidelity, and surpasses the comprehensive performance of existing numerical, analytical or other deep learning-based methods. It can reduce the modeling error by two orders of magnitude and ensure consistent high accuracy in both near-field and far-field scenarios. In addition, although the speed of the intelligent optical computing reverse design model is slightly inferior to the Fourier optics method model, it is 17,000 times faster than numerical simulation. In addition, its output is differentiable with respect to the input, allowing error back propagation. As an accurate and fast forward solver, the intelligent optical computing reverse design model can ensure high fidelity and efficiency of the design and is suitable for scenarios that require a large number of evaluations.

[0082] It should be noted that although the above embodiment only shows the design of a supersystem composed of a one-dimensional sub-wavelength structure, if the one-dimensional tensor in the intelligent optical computing reverse design model is extended to two dimensions, efficient and accurate reverse design of a two-dimensional sub-wavelength structure can be achieved. In addition, by increasing the number of channels in the tensor, each channel represents a different frequency component of the electromagnetic field, it is also possible to achieve reverse design of the device based on the frequency domain response.

[0083] In summary, the method provided by the embodiment of the present disclosure, by adopting the intelligent optical computing reverse design model, innovates a design paradigm that can ensure the balance of high efficiency and high fidelity, and has good scalability, showing greater flexibility than previous reverse design methods based on deep learning. More importantly, this flexible paradigm can be applied to the unprecedentedly challenging design of large-scale, high-degree-of-freedom and functionally complex devices, such as on-chip optical diffraction networks, thereby further promoting the development of computing super systems and paving the way for the development of large-scale, highly integrated and powerful optical intelligent computing.

[0084] In order to implement the above embodiments, the present disclosure also proposes an intelligent optical computing reverse design architecture.

[0085] like Figure 5 As shown, the intelligent optical computing reverse design architecture includes: A sample acquisition unit, used to acquire a sample data set, wherein the sample data set includes an input electromagnetic field sample; A parameter training unit, used to train the structural parameters of the intelligent optical computing device to be designed based on the sample data set and the intelligent optical computing reverse design model shown in the above embodiment, so as to obtain target structural parameters that meet the training requirements; The device construction unit is used to construct an intelligent optical computing device corresponding to the target structural parameters.

[0086] Optionally, the parameter training unit is used to train the structural parameters of the intelligent optical computing device to be designed, and when the target structural parameters that meet the training requirements are obtained, it is specifically used to: Inputting the structural parameters of the intelligent optical computing device to be designed and any input electromagnetic field sample in the sample data set into the intelligent optical computing reverse design model to obtain a predicted output electromagnetic field; Input the structural parameters and any input electromagnetic field sample into the numerical calculation solver to obtain the real output electromagnetic field corresponding to any input electromagnetic field sample; A second error value between the predicted output electromagnetic field and the actual output electromagnetic field is determined. If the second error value is greater than an error threshold, the structural parameters are updated until the second error value is no greater than the error threshold, thereby obtaining target structural parameters that meet the training requirements.

[0087] It should be noted that the aforementioned explanation of the embodiment of the intelligent optical computing reverse design method is also applicable to the intelligent optical computing reverse design architecture of this embodiment, which will not be repeated here.

[0088] In summary, the architecture provided in this embodiment, by adopting the intelligent optical computing reverse design model, innovates a design paradigm that can ensure the balance of high efficiency and high fidelity, and has good scalability, showing greater flexibility than previous reverse design methods based on deep learning. More importantly, this flexible paradigm can be applied to the unprecedentedly challenging design of large-scale, high-degree-of-freedom, and functionally complex devices, such as on-chip optical diffraction networks, thereby further promoting the development of computing super systems and paving the way for the development of large-scale, highly integrated, and powerful optical intelligent computing.

[0089] The collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in this disclosure shall comply with the relevant laws and regulations and shall not violate public order and good morals.

[0090] It should be noted that personal information from users should be collected for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. In addition, such collection / sharing should be carried out after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign the agreement / authorization including authorization of relevant user information before the user uses the function. In addition, any necessary steps should be taken to protect and safeguard access to such personal information data and ensure that others who have access to personal information data comply with its privacy policy and procedures.

[0091] The present disclosure anticipates providing implementation schemes for users to selectively block the use or access of personal information data. That is, the present disclosure anticipates providing hardware and / or software to prevent or block access to such personal information data. Once the personal information data is no longer needed, risks can be minimized by limiting data collection and deleting the data. In addition, when applicable, such personal information is de-identified to protect the privacy of the user.

[0092] The acquisition, transmission, storage, use, and processing of data in the technical solution disclosed in this disclosure are in compliance with the relevant provisions of national laws and regulations.

[0093] It should be noted that in the embodiments of the present disclosure, certain software, components, models and other existing solutions in the industry may be mentioned, which should be regarded as exemplary. Their purpose is only to illustrate the feasibility of implementing the technical solution of the present application, but it does not mean that the applicant has or will necessarily use the solution.

[0094] In the description of the aforementioned embodiments, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they contradict each other.

[0095] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present disclosure, "plurality" means at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0096] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present disclosure includes additional implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present disclosure belong.

[0097] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or otherwise processing in a suitable manner if necessary, and then stored in a computer memory.

[0098] It should be understood that the various parts of the present disclosure can be implemented in hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0099] A person skilled in the art may understand that all or part of the steps in the above-mentioned embodiment method may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0100] In addition, each functional unit in each embodiment of the present disclosure may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0101] The storage medium mentioned above may be a read-only memory, a disk or an optical disk, etc. Although the embodiments of the present disclosure have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations of the present disclosure. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present disclosure.

Claims

1. An intelligent optical computing reverse design model, characterized in that: include: An input partitioning module, for partitioning the input electromagnetic field into a set of input electromagnetic field segments based on the Huygens-Fresnel law and according to a total wavefront width of the input electromagnetic field and a wavefront width corresponding to each electromagnetic neural local solver in the set of electromagnetic neural local solvers; A wavefront prediction module, used for inputting each input electromagnetic field segment in the input electromagnetic field segment set and the structural parameters corresponding to the intelligent optical computing device to be designed into the electromagnetic neural local solver corresponding to the input electromagnetic field segment, so that the electromagnetic neural local solver predicts the output electromagnetic field segment according to the input electromagnetic field segment and the structural parameters to obtain the output electromagnetic field segment set; The output splicing module is used to splice all the output electromagnetic field segments in the output electromagnetic field segment set based on the Huygens-Fresnel law to obtain an output electromagnetic field under a global response, so as to adjust the structural parameters according to the output electromagnetic field.

2. The model according to claim 1, characterized in that The input division module is used to divide the input electromagnetic field into an input electromagnetic field segment set according to the total wavefront width of the input electromagnetic field and the wavefront width corresponding to each electromagnetic neural local solver in the electromagnetic neural local solver set, specifically for: The total wavefront width of the input electromagnetic field is divided based on the wavefront width corresponding to each electromagnetic neural local solver in the electromagnetic neural local solver set to obtain a division result; If the division result indicates that there is an input electromagnetic field segment with a width smaller than the wavefront width in the input electromagnetic field, zero padding is added to the end side and / or the end side of the input electromagnetic field until the width of each input electromagnetic field segment is the wavefront width, thereby obtaining an input electromagnetic field segment set, wherein the input electromagnetic field segment is regarded as a secondary wave generated by a central source point on the input electromagnetic field segment.

3. The model according to claim 1, characterized in that When the electromagnetic neural local solver is used to predict the output electromagnetic field segment according to the input electromagnetic field segment and the structural parameters, it is specifically used to: Performing complex linear connection on the input electromagnetic field segments to represent the non-physical information propagation process of the input electromagnetic field segments in the intelligent optical computing device to be designed, and obtaining the input electromagnetic field segments after non-physical propagation; Performing real number nonlinear connection on the structural parameters, and aligning information between the connected structural parameters and the input electromagnetic field segment after non-physical propagation, so as to represent the physical information propagation process of the input electromagnetic field segment after propagation in the intelligent optical computing device to be designed, and obtaining the input electromagnetic field segment after physical propagation; The input electromagnetic field segments after physical propagation are connected in a complex linear manner to represent the non-physical information propagation process of the input electromagnetic field segments after physical propagation in the intelligent optical computing device to be designed, and output electromagnetic field segments are obtained.

4. The model according to claim 3, characterized in that When the electromagnetic neural local solver is used for complex linear connection, it is specifically used for: Acquire a first weight matrix and a first intercept term used in performing complex linear connection, and perform complex linear connection based on the first weight matrix and the first intercept term, wherein the first weight matrix and the first intercept term are obtained by training an initial electromagnetic neural local solver.

5. The model according to claim 3, characterized in that When the electromagnetic neural local solver is used for real number nonlinear connection, it is specifically used for: A second weight matrix, a second intercept term and a nonlinear operation function used in performing real nonlinear connection are obtained, and a complex linear connection is performed based on the second weight matrix, the second intercept term and the nonlinear operation function, wherein the second weight matrix, the second intercept term and the nonlinear operation function are obtained by training an initial electromagnetic neural local solver.

6. The model according to claim 1, characterized in that Before inputting each input electromagnetic field segment in the input electromagnetic field segment set and the structural parameters corresponding to the intelligent optical computing device to be designed into the electromagnetic neural local solver corresponding to the input electromagnetic field segment, the wavefront prediction module is further used to: Acquire a training sample set, wherein each training sample in the training sample set includes an input electromagnetic field training sample and a structural parameter training sample; Performing gradient back propagation training on the initial electromagnetic neural local solver based on the training sample set until an electromagnetic neural local solver that meets the training requirements is obtained; Acquire a test sample set, wherein each of the test samples includes an input electromagnetic field test sample and a structural parameter test sample; The electromagnetic neural local solver is tested based on the test sample set. If the test result does not meet the test requirements, the electromagnetic neural local solver is re-trained by gradient back propagation until an electromagnetic neural local solver whose test result meets the test requirements is obtained.

7. The model according to claim 6, characterized in that When the wavefront prediction module is used to perform gradient back propagation training on the initial electromagnetic neural local solver based on the training sample set, it is specifically used to: In the process of performing gradient back-propagation training on the initial electromagnetic neural local solver based on the training sample set, any training sample in the training sample set is input into the initial electromagnetic neural local solver to obtain a predicted value corresponding to any training sample, and a first error value between the predicted value and a true value corresponding to any training sample is determined, and the network parameters of the initial electromagnetic neural local solver are updated based on the first error value to minimize the loss function.

8. An intelligent optical computing reverse design method, characterized in that: include: Acquire a sample data set, wherein the sample data set includes an input electromagnetic field sample; Based on the sample data set and the intelligent optical computing reverse design model according to any one of claims 1 to 7, training the structural parameters of the intelligent optical computing device to be designed to obtain target structural parameters that meet the training requirements; An intelligent optical computing device corresponding to the target structural parameters is constructed.

9. The method according to claim 8, characterized in that The step of training the structural parameters of the intelligent optical computing device to be designed to obtain target structural parameters that meet the training requirements includes: Inputting the structural parameters of the intelligent optical computing device to be designed and any input electromagnetic field sample in the sample data set into the intelligent optical computing reverse design model to obtain a predicted output electromagnetic field; Inputting the structural parameters and any input electromagnetic field sample into a numerical calculation solver to obtain a real output electromagnetic field corresponding to any input electromagnetic field sample; Determine a second error value between the predicted output electromagnetic field and the actual output electromagnetic field. If the second error value is greater than an error threshold, update the structural parameters until the second error value is no greater than the error threshold, thereby obtaining target structural parameters that meet training requirements.

10. An intelligent optical computing reverse design architecture, characterized in that: include: A sample acquisition unit, used to acquire a sample data set, wherein the sample data set includes an input electromagnetic field sample; A parameter training unit, configured to train the structural parameters of the intelligent optical computing device to be designed based on the sample data set and the intelligent optical computing reverse design model according to any one of claims 1 to 7, so as to obtain target structural parameters that meet the training requirements; The device construction unit is used to construct an intelligent optical computing device corresponding to the target structural parameters.

Images