Topological optical switch structure design method and system based on multi-network cascade model

Abstract

The present invention discloses a topological optical switch structure design method and system based on a multi-network cascade model, relating to the field of topological optical switch structure design. The topological optical switch structure design method based on the multi-network cascade model mainly includes: constructing a one-dimensional convolution inverse design network, an amplitude coupling ratio prediction network, a mathematical model based on the transfer matrix method, and a transmission spectrum forward prediction model; independently training each sub-model, cascading the trained sub-models so that the output of each network becomes the input of the next network to obtain a multi-network cascade model; training the multi-network cascade model as a whole to obtain an optimized one-dimensional convolution inverse design network; and obtaining target structural parameters using the optimized one-dimensional convolution inverse design network based on target transmission spectrum data. Implementing the topological optical switch structure design method and system based on the multi-network cascade model provided by the present invention can improve the accuracy of topological optical switch structure design.

Description

Technical Field

[0001] The present invention relates to the field of topological optical switch structure design, and more particularly to a topological optical switch structure design method and a topological optical switch structure design system based on a multi-network cascade model. Background Art

[0002] Optical switches are one of the important components for modern optical networks to develop towards higher speeds and larger capacities. Whether it is long-distance optical signal switching or on-chip optical interconnection, optical switches are key optical devices.

[0003] Topological optical switches, with their edge state properties, enable precise control of photon transmission, offering potential applications in optical communications, optical computing, and quantum information processing. In the field of topological photonics, the inverse design of the structural parameters of topological optical switches is a challenging task, requiring consideration of not only the functional implementation of photonic devices but also addressing the inherent non-uniqueness and multiple degrees of freedom inherent in the design process.

[0004] Traditional numerical simulation methods for simple topological optical switch structures often require significant time and computational resources, often requiring fine-tuning of the geometry and iterative simulations to gradually approach the target response. This process relies heavily on previous design experience, and due to limitations in computer hardware and simulation time, the final design may only result in a locally optimal structure.

[0005] Unlike traditional approaches based on physical models and strategies, deep learning (DL) is a data-driven approach. The photonics community has also benefited from the rapid development of deep learning technology, enabling intelligent reverse design of many photonic devices. Deep learning provides a new approach for solving forward prediction and reverse design problems related to spectral response. Well-trained deep learning models can directly map structural parameter designs to the optical properties of target photonic devices, and vice versa.

[0006] However, in the design of topological optical switches, determining optical topology requires considering the interaction between structural parameters and electromagnetic waves. The transmission spectrum response and photonic bandgap jointly determine the performance of topological optical switches. Currently, there is a lack of inverse design methods that meet the performance constraints of optical switches within this high degree of freedom. Furthermore, current deep learning-based inverse design of optical devices primarily focuses on external electromagnetic responses, such as transmission, reflection, and absorption spectra. There is limited research on the quantification and prediction of photonic band diagrams, and quantifying band diagrams into data that neural networks can learn from is a key issue.

[0007] In summary, the electromagnetic response in the existing inverse design of photonic devices corresponds to multiple structural parameters, and the few constraints used lead to large errors in the predicted parameters. The existing inverse design of micro-ring resonator (MRR) optical switches has problems of non-uniqueness and multiple degrees of freedom. Summary of the Invention

[0008] The object of the present invention is to provide a topological optical switch structure design method and a topological optical switch structure design system based on a multi-network cascade model, which can improve the accuracy of topological optical switch structure design.

[0009] The present invention provides a topological optical switch structure design method based on a multi-network cascade model, comprising the following steps:

[0010] S1: Using the parameter scanning method and the finite-difference time-domain method, a comprehensive data set is obtained. The comprehensive data set includes transmission spectrum data, structural parameter data, amplitude coupling ratio data, and photon band image data;

[0011] S2: Construct a one-dimensional convolution inverse design network, an amplitude coupling ratio prediction network, a mathematical model based on the transfer matrix method, and a transmission spectrum forward prediction model; the one-dimensional convolution inverse design network is used to predict structural parameter data based on transmission spectrum data; the amplitude coupling ratio prediction network is used to predict amplitude coupling ratio data based on structural parameter data; the mathematical model based on the transfer matrix method is used to calculate photon band image data based on structural parameter data and amplitude coupling ratio data; the transmission spectrum forward prediction model is used to predict transmission spectrum data based on photon band image data;

[0012] S3: Using the comprehensive dataset, the one-dimensional convolutional inverse design network, the amplitude coupling ratio prediction network, and the transmission spectrum forward prediction model are independently trained to obtain the trained one-dimensional convolutional inverse design network, the trained amplitude coupling ratio prediction network, and the trained transmission spectrum forward prediction model;

[0013] S4: Based on the trained one-dimensional convolutional inverse design network, the trained amplitude coupling ratio prediction network, the mathematical model based on the transfer matrix method, and the trained transmission spectrum forward prediction model, a multi-network cascade model is obtained. Based on the comprehensive dataset, the multi-network cascade model is trained as a whole using a loss function to obtain an optimized one-dimensional convolutional inverse design network.

[0014] S5: Based on the target transmission spectrum data, the target structure parameters are obtained using the optimized one-dimensional convolution inverse design network.

[0015] Furthermore, the one-dimensional convolution inverse design network of the topological optical switch structure design method based on the multi-network cascade model includes a one-dimensional convolutional neural network, which is used to learn the mapping relationship between the transmission spectrum and the structural parameters. The one-dimensional convolutional neural network includes a convolution layer and a fully connected layer; the amplitude coupling ratio prediction network takes the output of the one-dimensional convolution inverse design network as input and outputs the amplitude coupling ratio data. The amplitude coupling ratio prediction network is a deep fully connected neural network. The amplitude coupling ratio prediction network includes a 512×64×16×16×3 fully connected layer, and each neuron in the fully connected layer is connected to all neurons in the previous layer; the mathematical model based on the transfer matrix method receives the one-dimensional convolution inverse design network. The structural parameters output by the calculation network and the amplitude coupling ratio prediction network output the amplitude coupling ratio data, calculate and generate the photon band image; the transmission spectrum forward prediction model includes a photon band encoder and a band-transmission spectrum decoder; the photon band encoder is used to receive the photon band image of size 267×200 calculated by the transfer matrix method, and encode it into a one-dimensional band image code with a length of 8192; the photon band encoder includes a convolutional layer and a fully connected layer; the band-transmission spectrum decoder is used to receive the one-dimensional band image code of 8192 output by the photon band encoder, and decode it back to one-dimensional transmission spectrum data with a length of 501; the band-transmission spectrum decoder includes a convolutional layer and a fully connected layer.

[0016] Furthermore, the mathematical model based on the transmission matrix method of the topological optical switch structure design method based on the multi-network cascade model predicts the amplitude coupling ratio data of the network output according to the one-dimensional convolution inverse design of the structural parameters and amplitude coupling ratio of the network output, and uses the transmission matrix method and Bloch theorem to obtain the eigenvalue according to the eigenvalue equation, discards each eigenvalue that does not have a unit mode, and then scans the Bloch wave vector k x and phase, we get the photon band image, as shown in the formula:

[0017]

[0018]

[0019]

[0020]

[0021] Among them, M x M is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the x direction, y is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the y direction, k and τ are the amplitude coupling ratio and amplitude transmittance between the field ring and the connecting ring, respectively, and both satisfy |k|2 +|τ| 2 =1, Φ is the phase shift produced when the optical path is one quarter of the micro-ring circumference. det(S) is The determinant of k x is the Bloch wave vector, b' is the input amplitude at the coupling point between the adjacent connecting ring and the field ring, and d' is the output amplitude at the coupling point between the adjacent connecting ring and the field ring.

[0022] Furthermore, the loss function of the above-mentioned topological optical switch structure design method based on the multi-network cascade model is as follows:

[0023]

[0024] Among them, L δ (y,f(x)) is the loss function, y is the true parameter, f(x) is the predicted parameter, and δ is the parameter of Huber Loss.

[0025] The present invention also provides a topological optical switch structure design system, comprising the following modules:

[0026] A comprehensive data set acquisition module is configured to: obtain a comprehensive data set using a parameter scanning method and a finite-difference time-domain method, the comprehensive data set including transmission spectrum data, structural parameter data, amplitude coupling ratio data, and photon band image data;

[0027] A sub-model construction module is configured to: construct a one-dimensional convolution inverse design network, an amplitude coupling ratio prediction network, a mathematical model based on the transfer matrix method, and a transmission spectrum forward prediction model; the one-dimensional convolution inverse design network is used to predict structural parameter data based on transmission spectrum data; the amplitude coupling ratio prediction network is used to predict amplitude coupling ratio data based on structural parameter data; the mathematical model based on the transfer matrix method is used to calculate photon band image data based on structural parameter data and amplitude coupling ratio data; the transmission spectrum forward prediction model is used to predict transmission spectrum data based on photon band image data;

[0028] The sub-model training module is configured to: use the comprehensive data set to independently train the one-dimensional convolution inverse design network, the amplitude coupling ratio prediction network, and the transmission spectrum forward prediction model, thereby obtaining a trained one-dimensional convolution inverse design network, a trained amplitude coupling ratio prediction network, and a trained transmission spectrum forward prediction model;

[0029] The model training and optimization module is configured to: obtain a multi-network cascade model based on the trained one-dimensional convolutional inverse design network, the trained amplitude coupling ratio prediction network, the mathematical model based on the transfer matrix method, and the trained transmission spectrum forward prediction model; and perform overall training of the multi-network cascade model using a loss function based on a comprehensive data set to obtain an optimized one-dimensional convolutional inverse design network.

[0030] The structural inverse design module is configured as follows: according to the target transmission spectrum data, the optimized one-dimensional convolution inverse design network is used to obtain the target structural parameters.

[0031] Furthermore, the one-dimensional convolution inverse design network of the above-mentioned topological optical switch structure design system includes a one-dimensional convolutional neural network, which is used to learn the mapping relationship between the transmission spectrum and the structural parameters. The one-dimensional convolutional neural network includes a convolution layer and a fully connected layer; the amplitude coupling ratio prediction network takes the output of the one-dimensional convolution inverse design network as input and outputs amplitude coupling ratio data. The amplitude coupling ratio prediction network is a deep fully connected neural network. The amplitude coupling ratio prediction network includes a 512×64×16×16×3 fully connected layer, and each neuron in the fully connected layer is connected to all neurons in the previous layer; the mathematical model based on the transfer matrix method receives the output of the one-dimensional convolution inverse design network The structural parameters of the network and the amplitude coupling ratio prediction network output the amplitude coupling ratio data to calculate and generate the photon band image; the transmission spectrum forward prediction model includes a photon band encoder and a band-transmission spectrum decoder; the photon band encoder is used to receive the photon band image of size 267×200 calculated by the transfer matrix method, and encode it into a one-dimensional band image code with a length of 8192; the photon band encoder includes a convolutional layer and a fully connected layer; the band-transmission spectrum decoder is used to receive the one-dimensional band image code of 8192 output by the photon band encoder, and decode it back to one-dimensional transmission spectrum data with a length of 501; the band-transmission spectrum decoder includes a convolutional layer and a fully connected layer.

[0032] Furthermore, the mathematical model of the topological optical switch structure design system based on the transfer matrix method predicts the amplitude coupling ratio data of the network output according to the structural parameters and amplitude coupling ratio of the one-dimensional convolution inverse design network output, and uses the transfer matrix method and Bloch theorem to obtain the eigenvalue according to the eigenvalue equation, discards each eigenvalue that does not have a unit mode, and then scans the Bloch wave vector k x and phase, we get the photon band image, as shown in the formula:

[0033]

[0034]

[0035]

[0036]

[0037] Among them, M x M is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the x direction, yis the transmission matrix representing the coupling relationship of the microring resonator coupling array in the y direction, k and τ are the amplitude coupling ratio and amplitude transmittance between the field ring and the connecting ring, respectively, and both satisfy |k| 2 +|τ| 2 =1, Φ is the phase shift produced when the optical path is one quarter of the micro-ring circumference. det(S) is The determinant of k x is the Bloch wave vector, b' is the input amplitude at the coupling point between the adjacent connecting ring and the field ring, and d' is the output amplitude at the coupling point between the adjacent connecting ring and the field ring.

[0038] Furthermore, the loss function of the above topological optical switch structure design system is as follows:

[0039]

[0040] Among them, L δ (y,f(x)) is the loss function, y is the true parameter, f(x) is the predicted parameter, and δ is the parameter of Huber Loss.

[0041] The implementation of the topological optical switch structure design method and topological optical switch structure design system based on the multi-network cascade model provided by the present invention has the following beneficial effects:

[0042] The present invention draws on the idea of ​​cascade model and makes innovative improvements based on the traditional deep learning cascade network. It uses advanced deep learning architecture to connect multiple sub-models in a specific order and manner to form a more powerful and complex overall model. Specifically, the present invention proposes the use of a multi-network cascaded deep learning model for structural reverse design in combination with the design goals of topological optical switch performance parameters. The model consists of a reverse design model, an amplitude coupling ratio prediction model and a transmission spectrum forward prediction model. Its training and optimization process integrates multi-dimensional data sets such as transmission spectrum, structural parameters, amplitude coupling ratio and photon energy band, so that the entire model can learn and process more complex data, so that it can combine multi-degree-of-freedom optical response data to achieve more accurate and efficient supervised learning, thereby obtaining optical characteristics with more stable transmission and better switching performance, solving the non-uniqueness problem and multi-degree-of-freedom problem in the structural design of topological optical switches. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0044] Figure 1 It is a flow chart of a topological optical switch structure design method based on a multi-network cascade model provided by the present invention;

[0045] Figure 2 This is a schematic diagram of the microring array structure provided by the present invention;

[0046] Figure 3 It is a schematic diagram of a multi-network cascade model provided by the present invention;

[0047] Figure 4 This is a flow chart of a multi-network cascade model of a topological optical switch provided by the present invention;

[0048] Figure 5 This is the overall training flow chart of the cascade model provided by the present invention;

[0049] Figure 6 This is a comparison chart of the overall training loss of the cascade model provided by the present invention;

[0050] Figure 7 This is a comparison diagram of the predicted transmission spectrum dB graph provided by the present invention;

[0051] Figure 8 It is a comparison chart of the transmission spectrum simulation results of the structural parameters predicted by the direct inverse model provided by the present invention and the inverse model after multi-network cascade. DETAILED DESCRIPTION

[0052] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0053] Figure 1 A schematic diagram of a topological optical switch structure design method based on a multi-network cascade model of this embodiment is shown. In this embodiment, the topological optical switch structure design method based on a multi-network cascade model includes the following steps:

[0054] S1: Using the parameter scanning method and the finite-difference time-domain method, a comprehensive data set is obtained. The comprehensive data set includes transmission spectrum data, structural parameter data, amplitude coupling ratio data, and photon band image data;

[0055] S2: Construct a one-dimensional convolution inverse design network, an amplitude coupling ratio prediction network, a mathematical model based on the transfer matrix method, and a transmission spectrum forward prediction model; the one-dimensional convolution inverse design network is used to predict structural parameter data based on transmission spectrum data; the amplitude coupling ratio prediction network is used to predict amplitude coupling ratio data based on structural parameter data; the mathematical model based on the transfer matrix method is used to calculate photon band image data based on structural parameter data and amplitude coupling ratio data; the transmission spectrum forward prediction model is used to predict transmission spectrum data based on photon band image data;

[0056] Specifically, the one-dimensional convolution inverse design network of the topological optical switch structure design method based on the multi-network cascade model includes a one-dimensional convolutional neural network, which is used to learn the mapping relationship between the transmission spectrum and the structural parameters. The one-dimensional convolutional neural network includes a convolution layer and a fully connected layer; the amplitude coupling ratio prediction network takes the output of the one-dimensional convolution inverse design network as input and outputs the amplitude coupling ratio data. The amplitude coupling ratio prediction network is a deep fully connected neural network. The amplitude coupling ratio prediction network includes a 512×64×16×16×3 fully connected layer, and each neuron in the fully connected layer is connected to all neurons in the previous layer; the mathematical model based on the transfer matrix method receives the one-dimensional convolution inverse design. The network outputs structural parameters and amplitude coupling ratio prediction network output amplitude coupling ratio data, calculates and generates the photon band image; the transmission spectrum forward prediction model includes a photon band encoder and a band-transmission spectrum decoder; the photon band encoder is used to receive the photon band image of size 267×200 calculated by the transfer matrix method, and encode it into a one-dimensional band image code with a length of 8192; the photon band encoder is constructed by a convolutional layer and a fully connected layer; the band-transmission spectrum decoder is used to receive the one-dimensional band image code of 8192 output by the photon band encoder, and decode it back to a one-dimensional transmission spectrum data with a length of 501; the band-transmission spectrum decoder includes a convolutional layer and a fully connected layer;

[0057] Specifically, the mathematical model based on the transmission matrix method of the topological optical switch structure design method based on the multi-network cascade model predicts the amplitude coupling ratio data of the network output according to the one-dimensional convolution inverse design network output structural parameters and amplitude coupling ratio, and uses the transmission matrix method and Bloch theorem to obtain the eigenvalue according to the eigenvalue equation, discards each eigenvalue that does not have a unit mode, and then scans the Bloch wave vector k x and phase, we get the photon band image, as shown in the formula:

[0058]

[0059]

[0060]

[0061]

[0062] Among them, M x M is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the x direction, y is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the y direction, k and τ are the amplitude coupling ratio and amplitude transmittance between the field ring and the connecting ring, respectively, and both satisfy |k| 2+|τ| 2 =1, Φ is the phase shift produced when the optical path is one quarter of the micro-ring circumference. det(S) is The determinant of k x is the Bloch wave vector, b' is the input amplitude at the coupling point between the adjacent connecting ring and the field ring, and d' is the output amplitude at the coupling point between the adjacent connecting ring and the field ring;

[0063] S3: Using the comprehensive dataset, the one-dimensional convolutional inverse design network, the amplitude coupling ratio prediction network, and the transmission spectrum forward prediction model are independently trained to obtain the trained one-dimensional convolutional inverse design network, the trained amplitude coupling ratio prediction network, and the trained transmission spectrum forward prediction model;

[0064] S4: Based on the trained one-dimensional convolutional inverse design network, the trained amplitude coupling ratio prediction network, the mathematical model based on the transfer matrix method, and the trained transmission spectrum forward prediction model, a multi-network cascade model is obtained; based on the comprehensive data set, the multi-network cascade model is trained as a whole using the loss function to obtain the optimized one-dimensional convolutional inverse design network;

[0065] Specifically, the loss function of the above-mentioned topological optical switch structure design method based on the multi-network cascade model is as follows:

[0066]

[0067] Among them, L δ (y,f(x)) is the loss function, y is the true parameter, f(x) is the predicted parameter, and δ is the parameter of Huber Loss;

[0068] S5: Based on the target transmission spectrum data, the target structure parameters are obtained using the optimized one-dimensional convolution inverse design network.

[0069] The present invention also provides a topological optical switch structure design system, comprising the following modules:

[0070] A comprehensive data set acquisition module is configured to: obtain a comprehensive data set using a parameter scanning method and a finite-difference time-domain method, the comprehensive data set including transmission spectrum data, structural parameter data, amplitude coupling ratio data, and photon band image data;

[0071] A sub-model construction module is configured to: construct a one-dimensional convolution inverse design network, an amplitude coupling ratio prediction network, a mathematical model based on the transfer matrix method, and a transmission spectrum forward prediction model; the one-dimensional convolution inverse design network is used to predict structural parameter data based on transmission spectrum data; the amplitude coupling ratio prediction network is used to predict amplitude coupling ratio data based on structural parameter data; the mathematical model based on the transfer matrix method is used to calculate photon band image data based on structural parameter data and amplitude coupling ratio data; the transmission spectrum forward prediction model is used to predict transmission spectrum data based on photon band image data;

[0072] Specifically, the one-dimensional convolution inverse design network of the topological optical switch structure design system includes a one-dimensional convolutional neural network, which is used to learn the mapping relationship between the transmission spectrum and the structural parameters. The one-dimensional convolutional neural network includes a convolution layer and a fully connected layer. The amplitude coupling ratio prediction network takes the output of the one-dimensional convolution inverse design network as input and outputs the amplitude coupling ratio data. The amplitude coupling ratio prediction network is a deep fully connected neural network. The amplitude coupling ratio prediction network includes a 512×64×16×16×3 fully connected layer. Each neuron in the fully connected layer is connected to all neurons in the previous layer. The mathematical model based on the transfer matrix method receives the output of the one-dimensional convolution inverse design network. The amplitude coupling ratio data output by the structural parameter and amplitude coupling ratio prediction network is used to calculate and generate a photon band image; the transmission spectrum forward prediction model includes a photon band encoder and a band-transmission spectrum decoder; the photon band encoder is used to receive the photon band image of size 267×200 calculated by the transfer matrix method and encode it into a one-dimensional band image code of length 8192; the photon band encoder is constructed by convolutional layers and fully connected layers; the band-transmission spectrum decoder is used to receive the one-dimensional band image code of 8192 output by the photon band encoder and decode it back into one-dimensional transmission spectrum data of length 501; the band-transmission spectrum decoder includes convolutional layers and fully connected layers;

[0073] Specifically, the mathematical model based on the transfer matrix method of the topological optical switch structure design system predicts the amplitude coupling ratio data of the network output according to the structural parameters and amplitude coupling ratio of the one-dimensional convolution inverse design network output, and uses the transfer matrix method and Bloch theorem to obtain the eigenvalue according to the eigenvalue equation, discards each eigenvalue that does not have a unit mode, and then scans the Bloch wave vector k x and phase, we get the photon band image, as shown in the formula:

[0074]

[0075]

[0076]

[0077]

[0078] Among them, M x M is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the x direction, y is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the y direction, k and τ are the amplitude coupling ratio and amplitude transmittance between the field ring and the connecting ring, respectively, and both satisfy |k| 2 +|τ| 2 =1, Φ is the phase shift produced when the optical path is one quarter of the micro-ring circumference. det(S) is The determinant of k x is the Bloch wave vector, b' is the input amplitude at the coupling point between the adjacent connecting ring and the field ring, and d' is the output amplitude at the coupling point between the adjacent connecting ring and the field ring;

[0079] The sub-model training module is configured to: use the comprehensive data set to independently train the one-dimensional convolution inverse design network, the amplitude coupling ratio prediction network, and the transmission spectrum forward prediction model, thereby obtaining a trained one-dimensional convolution inverse design network, a trained amplitude coupling ratio prediction network, and a trained transmission spectrum forward prediction model;

[0080] The model training and optimization module is configured to: obtain a multi-network cascade model based on the trained one-dimensional convolution inverse design network, the trained amplitude coupling ratio prediction network, the mathematical model based on the transfer matrix method, and the trained transmission spectrum forward prediction model; and perform overall training of the multi-network cascade model using a loss function based on a comprehensive data set to obtain an optimized one-dimensional convolution inverse design network.

[0081] Specifically, the loss function of the above topological optical switch structure design system is as follows:

[0082]

[0083] Among them, L δ (y,f(x)) is the loss function, y is the true parameter, f(x) is the predicted parameter, and δ is the parameter of Huber Loss;

[0084] The structural inverse design module is configured as follows: according to the target transmission spectrum data, the optimized one-dimensional convolution inverse design network is used to obtain the target structural parameters.

[0085] In some embodiments, the above-mentioned topological optical switch structure design method based on the multi-network cascade model can also be implemented in the following manner.

[0086] In this embodiment, the structural model of the topological microring array used is as follows: Figure 2 As shown, it consists of a 3×3 microring array and input and output waveguides. The black rings in the figure are field rings, and the gray rings are coupling rings. One input waveguide is connected to the middle field ring on the left end of the array, and one output waveguide is symmetrical to it, and the other is connected to the field ring at the top right of the array. The area within the frame is the modulation area, in which the microrings are thermo-optically modulated. The structural parameters are the field ring radius R1, the coupling ring radius R2, and the coupling spacing g between the field ring and the coupling ring.

[0087] Specifically, the structural parameters of the field ring radius R1 and the coupling ring radius R2 are uniformly generated within the range of 1.3μm to 1.9μm with an accuracy of 0.02μm, and the coupling spacing g between the field ring and the coupling ring is uniformly generated within the range of 0.02μm to 0.12μm with an accuracy of 0.01μm;

[0088] Specifically, a parameter sweep method was used to generate 10,571 simulated transmission spectra in a two-dimensional FDTD model. Each transmission spectrum consisted of 501 data points, and the transmission spectrum wavelength range was 1500-1600 nm. The ratio of the training set to the test set was 8:2.

[0089] Specifically, the established multi-network cascade model is as follows Figure 3 As shown in the figure, it consists of a one-dimensional convolution inverse design model, an amplitude coupling ratio prediction model, a photon energy band calculation module based on the transfer matrix method, and a transmission spectrum forward prediction model. The input and output of the overall model are both transmission spectrum data. The loss function is calculated based on the error between the predicted transmission spectrum and the true transmission spectrum and backpropagation is performed. This ensures a one-to-one correspondence between the predicted transmission spectrum and the true transmission spectrum, eliminating the impact of non-uniqueness on prediction accuracy. In addition, the multi-network cascade model integrates multi-dimensional data sets such as transmission spectrum, structural parameters, amplitude coupling ratio, and photon energy band, which can further improve the accuracy of the model.

[0090] Specifically, in order to enhance the expressive power of the network, the weights of each layer of the network are weighted and summed to complete the nonlinear transformation. The activation function uses the ReLU activation function, which is the most commonly used activation function in deep neural networks. It has the advantages of fast operation speed and anti-gradient disappearance. Its expression is:

[0091] Relu(x)=max(0,x) (1)

[0092] Specifically, the loss function is used to measure the difference between the model output and the true value. Huber Loss is a loss function used for regression problems. It combines the advantages of mean square error (MSE) and mean absolute error (MAE). This example uses it to calculate the prediction error of all models. Its definition is as follows:

[0093]

[0094] Where y is the true parameter, f(x) is the predicted parameter, and δ is the parameter of Huber Loss. When the prediction deviation is less than or equal to δ, Huber Loss uses MSE to reduce the model's sensitivity to small deviations. When the prediction deviation is greater than δ, it switches to MAE to reduce sensitivity to large deviations (such as outliers). It is a loss function with parameters used to solve regression problems.

[0095] Specifically, to evaluate the reliability of the model, this embodiment will calculate the mean relative error (MRE) and the accuracy of the network; MRE is the average value of the relative difference between the predicted value and the actual value, and its expression is:

[0096]

[0097] Where n is the sample size, y i is the true value, is the predicted value.

[0098] In contrast, the concept of accuracy is used. In this embodiment, for regression problems using minimum-maximum normalization, the accuracy is calculated as follows:

[0099]

[0100] Where N is the sample size, y i is the true value, is the predicted value, and the latter two are normalized values.

[0101] Specifically, the forward prediction model consists of an energy band image encoder and an energy band-transmission spectrum decoder. The coupling ratio prediction model and the transmission spectrum forward prediction model transmit data via photon energy bands. The specific architecture of the above model is as follows:

[0102] One-dimensional convolutional inverse design network: This network takes as input one-dimensional transmission spectrum data of length 501 and outputs one-dimensional structural parameter combination data of length 3. This network learns the complex mapping relationship between transmission spectrum and structural parameters through a one-dimensional convolutional neural network, thus realizing the inverse design function.

[0103] Specifically, the Pytorch framework in Python was used to build and train the model, employing a 1D-CNN consisting of convolutional and fully connected layers. First, the transmission spectrum data of length 501 was input, passed through four one-dimensional convolutional layers and a pooling layer, accelerated by batchnorm1d, and then sent to the fully connected layer for processing, ultimately obtaining a structural parameter size of 3.

[0104] Amplitude Coupling Ratio Prediction Network: This network takes as input the output of a one-dimensional convolutional inverse design network, i.e., one-dimensional structural parameter combination data of length 3, and generates amplitude coupling ratio data of length 1. This network builds a model using a deep fully connected neural network to accurately predict the required coupling ratio.

[0105] Specifically, a fully connected layer of 512×64×16×16×3 is used, where each neuron is connected to all neurons in the previous layer. Through these linear layers, the model gradually extracts key features from the data and ultimately obtains the coupling ratio k.

[0106] A MATLAB mathematical model based on the transfer matrix method is introduced into the model chain after the amplitude coupling ratio prediction network. This model receives the structural parameters and amplitude coupling ratio output by the aforementioned network and calculates and generates a photon band image. This mathematical model converts the prediction results of the aforementioned neural network into a photon band image, which provides input for the subsequent encoder.

[0107] Specifically, the transmission matrices representing the coupling relationship of the optical microring array in the x and y directions can be expressed as:

[0108]

[0109]

[0110] The above formula can be used to solve the projected energy band of periodic lattice, and can also be used to calculate the projected energy band of finite optical microring array;

[0111] Combining formulas (5), (6) and Bloch's theorem, we can simplify and obtain x and M y The eigenvalues ​​of the underlying periodic lattice are expressed as:

[0112]

[0113] Find the eigenvalue of formula (7) and transform the eigenvalue equation into:

[0114]

[0115] Discard each eigenvalue that does not have a unit modulus, and then scan k x The projected band diagram of the underlying periodic lattice can be obtained by using the phase;

[0116] Photon Band Encoder: This encoder receives a 267×200-pixel photon band image calculated by the transfer matrix method and encodes it into a one-dimensional band image code with a length of 8192. The encoder is designed to generate an efficient intermediate feature representation that relates the complex band image data to the transmission spectrum data.

[0117] Specifically, the band image encoder model is primarily constructed from convolutional and fully connected layers, which convert a two-dimensional band image into a one-dimensional code. The initial band image is 267×200 pixels in size. Four convolutional and pooling layers reduce the image size to 50×66 pixels, which is then converted to a one-dimensional code of 8192 pixels via a flattening layer. The flattening layer arranges the pixels in a one-dimensional order for subsequent fully connected layers to process. Next, the one-dimensional code is gradually reduced in dimension via four linear layers, ultimately resulting in a one-dimensional code of size 501 pixels.

[0118] Energy band-transmission spectrum decoder: Finally, the decoder receives the output of the energy band-transmission spectrum encoder, which is a one-dimensional energy band image code with a length of 8192, and decodes it back into a one-dimensional transmission spectrum data with a length of 501. The purpose of the decoder is to reconstruct the original transmission spectrum information, completing the data conversion process of the entire model.

[0119] The specific multi-network cascade model process is as follows Figure 4 As shown in the figure, the training process is divided into two stages. First, each sub-model is trained independently, and then these sub-models are integrated to train the entire cascade model. In the initial stage, a comprehensive data set including transmission spectrum, structural parameters, amplitude coupling ratio and photon energy band data needs to be generated, and necessary pre-processing operations are performed on the transmission spectrum and photon energy band data. Next, each sub-model is trained. The inverse design model is responsible for mapping the transmission spectrum of the topological optical switch to its corresponding structural parameters. The amplitude coupling ratio prediction model outputs the corresponding amplitude coupling ratio according to the input structural parameters. The transfer matrix method calculation module is based on The photon energy band is calculated and output based on the structural parameters and amplitude coupling ratio; the forward prediction model uses the photon energy band as input and predicts the corresponding transmission spectrum; at this stage, all trained sub-models will be properly saved; then, the trained sub-models will be cascaded to train the overall cascade model. During the training of the overall cascade model, the inverse design model will undergo a second training to optimize performance; after training, the overall model and the inverse design model will be saved; by loading the saved inverse design model file and inputting the target transmission spectrum, the structural parameters that meet the requirements of the ideal optical properties can be quickly predicted;

[0120] Finally, the predicted structural parameters are input into the FDTD simulation software to generate the transmission spectrum, which is then compared with the original prediction results to evaluate the accuracy and reliability of the model.

[0121] Specifically, the overall training process of the cascade model is as follows: Figure 5As shown in the figure, by directly calling each sub-model file, the cascade model is constructed according to the input-output relationship of each model. Among them, the link between the amplitude coupling ratio prediction model and the forward model requires the calculation of the photon energy band through MATLAB. This step is achieved by using Python to call the transfer matrix method to generate the MATLAB file of the photon energy band. In this way, the cascade model is constructed.

[0122] Specifically, during the overall training process, the number of MATLAB parallel processes was set to 20, and 20 independent computing processes were simultaneously launched to perform the photon energy band calculation task, which significantly improved the processing speed of computationally intensive tasks. The loss function of the overall model calculated the Huber Loss between the input transmission spectrum and the predicted transmission spectrum, and passed the error to the Adam optimizer to optimize the encoder parameters to ensure stable convergence of the training process. During the training process, the learning rate was reduced from the original 10E-4 to 10E-5. After multiple iterative training, the model continuously adjusted its parameters until the error in the predicted transmission spectrum stabilized, achieving optimized model performance.

[0123] exist Figure 6 (a) shows the overall training loss change of the multi-network cascade model, where the dark curve depicts the error on the training set, while the light curve reflects the performance on the test set. After 500 iterations of rigorous training, the loss value of the cascaded overall model on the test set dropped significantly to 0.00943. At the same time, the prediction accuracy of the model also increased from 89.3% to 91.3%. The training loss of the sub-reverse design model is shown in Figure 2. Figure 6 As shown in (b), the loss value on the test set is reduced to 0.00026, and the accuracy is improved from 96.6% to 98.4%. Table 1 shows the comparison results of the network prediction accuracy.

[0124] Table 1 Error comparison before and after overall training of multi-network cascade model

[0125] Training methods Overall cascade model prediction accuracy Prediction accuracy of the reverse design model Not cascaded 89.3％ 96.6％ After cascading 91.3％ 98.4％

[0126] Figure 7 The transmission spectrum dB plot predicted by model training is shown, where the dark solid line represents the true transmission spectrum, and the light dashed line represents the predicted transmission spectrum. It can be observed that the prediction overlap is higher in flat areas of the transmission spectrum, and the prediction effect is also improved in steep areas of the transmission spectrum.

[0127] FDTD simulations validate the training results of the cascaded overall model, allowing for a more accurate assessment of model performance. To ensure the rationality of the results, this study first selected a transmission spectrum from the test dataset and used both the direct inverse model and the fully trained sub-inverse model to predict structural parameters for this spectrum. Table 2 shows the structural parameter predictions obtained using the two models.

[0128] Table 2 Structural parameter prediction results

[0129]

[0130] The training effect of the cascaded overall model is verified by FDTD simulation, which can more accurately evaluate the model performance. To ensure the rationality of the results, a transmission spectrum is first selected from the test data set, and the structure parameters of this spectrum are predicted using the direct inverse model and the sub-inverse model after overall training. Table 2 shows the structural parameter results and average relative errors predicted by the two models. It can be clearly seen that the error of the sub-inverse model after overall training is lower than that of the direct inverse model, and is controlled within 0.25E-2. Subsequently, the predicted structural parameter combinations are substituted into the FDTD simulation to obtain the transmission spectrum. The comparison of the transmission spectrum simulation results is shown in Table 2. Figure 8 As shown in the figure, the dark solid line, dark dashed line, and light dashed line represent the true transmission spectrum, the simulated transmission spectrum of the structural parameters predicted by the direct inverse model, and the simulated transmission spectrum of the structural parameters predicted by the sub-inverse model after overall training, respectively. By observing the transmission spectrum comparison chart, it can be found that the transmission spectrum obtained in the simulation of the structural parameters predicted by the sub-inverse model after overall training is more consistent with the true transmission spectrum. For topological optical switch structures, even slight deviations in the structural parameters will cause significant deviations in the transmission spectrum, so a more accurate deep learning model is needed to ensure the accuracy and reliability of the prediction results.

[0131] This proves that the use of a multi-network cascade model can effectively eliminate the interference caused by non-uniqueness problems and provide effective auxiliary training for the sub-inverse model, making the sub-inverse model more accurate.

[0132] Based on the principles of deep learning and transfer matrix method, this paper proposes a multi-network cascade model. By cascading, the advantages of multiple models are integrated into a unified framework to achieve more accurate predictions. It is suitable for solving the non-uniqueness and multi-degree-of-freedom problems in the reverse design of topological optical switches. It mainly includes the design and training of a one-dimensional convolution inverse design model, an amplitude coupling ratio prediction model, a forward prediction model and an overall model. In the process of constructing the cascade model, each sub-model is first trained to ensure that each sub-model achieves a high accuracy rate on the training set. Then, these sub-models are cascaded according to the design order of the cascade network, so that the output of each network becomes the next network. The input of the multi-network cascade model is a network; the photon energy band used to connect the amplitude coupling ratio prediction model and the forward prediction model is calculated using the transfer matrix method; the forward prediction model is trained using an encoder network, which converts the photon energy band and transmission spectrum, which are difficult to map, into network coding for indirect training, thereby improving the accuracy of the forward prediction model; finally, the multi-network cascade model is trained as a whole and verified by simulation in FDTD. The simulation results show that there is a high degree of consistency between the spectrum predicted by the model and the original spectrum; in addition, after training the multi-network cascade model, the accuracy of the trained cascade model is 91.3%, of which the accuracy of the inverse design model is 98.4%.

[0133] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A topological optical switch structure design method based on a multi-network cascade model, characterized in that: The following steps are involved: S1: using a parameter scanning method and a finite-difference time-domain method to obtain a comprehensive data set, wherein the comprehensive data set includes transmission spectrum data, structural parameter data, amplitude coupling ratio data, and photon band image data; S2: Constructing a one-dimensional convolution inverse design network, an amplitude coupling ratio prediction network, a mathematical model based on the transfer matrix method, and a transmission spectrum forward prediction model; the one-dimensional convolution inverse design network is used to predict structural parameter data based on transmission spectrum data; The amplitude coupling ratio prediction network is used to predict the amplitude coupling ratio data based on the structural parameter data; the mathematical model based on the transfer matrix method is used to calculate the photon band image data based on the structural parameter data and the amplitude coupling ratio data; the transmission spectrum forward prediction model is used to predict the transmission spectrum data based on the photon band image data; S3: using the comprehensive data set, independently training the one-dimensional convolution inverse design network, the amplitude coupling ratio prediction network, and the transmission spectrum forward prediction model to obtain a trained one-dimensional convolution inverse design network, a trained amplitude coupling ratio prediction network, and a trained transmission spectrum forward prediction model; S4: obtaining a multi-network cascade model based on the trained one-dimensional convolution inverse design network, the trained amplitude coupling ratio prediction network, the mathematical model based on the transfer matrix method, and the trained transmission spectrum forward prediction model; and performing overall training on the multi-network cascade model using a loss function based on the comprehensive data set to obtain an optimized one-dimensional convolution inverse design network. S5: According to the target transmission spectrum data, the target structure parameters are obtained by using the optimized one-dimensional convolution inverse design network.

2. The method for designing a topological optical switch structure based on a multi-network cascade model according to claim 1, characterized in that: The one-dimensional convolution inverse design network includes a one-dimensional convolutional neural network, which is used to learn the mapping relationship between the transmission spectrum and the structural parameters. The one-dimensional convolutional neural network includes a convolution layer and a fully connected layer. The amplitude coupling ratio prediction network takes the output of the one-dimensional convolution inverse design network as input and outputs amplitude coupling ratio data. The amplitude coupling ratio prediction network is a deep fully connected neural network. The amplitude coupling ratio prediction network includes a 512×64×16×16×3 fully connected layer, and each neuron in the fully connected layer is connected to all neurons in the previous layer. The mathematical model based on the transfer matrix method receives the structural parameters and amplitude coupling ratio data output by the one-dimensional convolution inverse design network. The amplitude coupling ratio prediction network outputs the amplitude coupling ratio data to calculate and generate a photon band image; the transmission spectrum forward prediction model includes a photon band encoder and a band-transmission spectrum decoder; the photon band encoder is used to receive the photon band image of size 267×200 calculated by the transfer matrix method, and encode it into a one-dimensional band image code with a length of 8192; the photon band encoder includes a convolutional layer and a fully connected layer; the band-transmission spectrum decoder is used to receive the one-dimensional band image code of 8192 output by the photon band encoder, and decode it back into one-dimensional transmission spectrum data with a length of 501; the band-transmission spectrum decoder includes a convolutional layer and a fully connected layer.

3. The method for designing a topological optical switch structure based on a multi-network cascade model according to claim 1, characterized in that: The mathematical model based on the transfer matrix method predicts the amplitude coupling ratio data of the network output according to the structural parameters and amplitude coupling ratio output by the one-dimensional convolution inverse design network, and uses the transfer matrix method and Bloch theorem to obtain the eigenvalue according to the eigenvalue equation, discards each eigenvalue that does not have a unit mode, and then scans the Bloch wave vector k x and phase, we get the photon band image, as shown in the formula: Among them, M x M is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the x direction, y is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the y direction, k and τ are the amplitude coupling ratio and amplitude transmittance between the field ring and the connecting ring, respectively, and both satisfy |k| 2 +|τ| 2 =1, Φ is the phase shift produced when the optical path is one quarter of the micro-ring circumference; det(S) is The determinant of k x is the Bloch wave vector, b' is the input amplitude at the coupling point between the adjacent connecting ring and the field ring, and d' is the output amplitude at the coupling point between the adjacent connecting ring and the field ring.

4. The method for designing a topological optical switch structure based on a multi-network cascade model according to claim 1, characterized in that: The loss function is as follows: Among them, L δ (y,f(x)) is the loss function, y is the true parameter, f(x) is the predicted parameter, and δ is the parameter of Huber Loss.

5. A topological optical switch structure design system, characterized in that: The topological optical switch structure design system includes the following modules: a comprehensive data set acquisition module configured to: obtain a comprehensive data set using a parameter scanning method and a finite-difference time-domain method, wherein the comprehensive data set includes transmission spectrum data, structural parameter data, amplitude coupling ratio data, and photon band image data; A sub-model construction module is configured to: construct a one-dimensional convolution inverse design network, an amplitude coupling ratio prediction network, a mathematical model based on the transfer matrix method, and a transmission spectrum forward prediction model; the one-dimensional convolution inverse design network is used to predict structural parameter data based on transmission spectrum data; The amplitude coupling ratio prediction network is used to predict the amplitude coupling ratio data based on the structural parameter data; the mathematical model based on the transfer matrix method is used to calculate the photon band image data based on the structural parameter data and the amplitude coupling ratio data; the transmission spectrum forward prediction model is used to predict the transmission spectrum data based on the photon band image data; a sub-model training module configured to: independently train the one-dimensional convolution inverse design network, the amplitude coupling ratio prediction network, and the transmission spectrum forward prediction model using the comprehensive data set, thereby obtaining a trained one-dimensional convolution inverse design network, a trained amplitude coupling ratio prediction network, and a trained transmission spectrum forward prediction model; A model training and optimization module is configured to obtain a multi-network cascade model based on the trained one-dimensional convolution inverse design network, the trained amplitude coupling ratio prediction network, the mathematical model based on the transfer matrix method, and the trained transmission spectrum forward prediction model; According to the comprehensive data set, the multi-network cascade model is trained as a whole using a loss function to obtain an optimized one-dimensional convolutional inverse design network; The structural inverse design module is configured to obtain target structural parameters based on target transmission spectrum data using the optimized one-dimensional convolution inverse design network.

6. A topological optical switch structure design system according to claim 5, characterized in that: The one-dimensional convolution inverse design network includes a one-dimensional convolutional neural network, which is used to learn the mapping relationship between the transmission spectrum and the structural parameters. The one-dimensional convolutional neural network includes a convolution layer and a fully connected layer. The amplitude coupling ratio prediction network takes the output of the one-dimensional convolution inverse design network as input and outputs amplitude coupling ratio data. The amplitude coupling ratio prediction network is a deep fully connected neural network. The amplitude coupling ratio prediction network includes a 512×64×16×16×3 fully connected layer, and each neuron in the fully connected layer is connected to all neurons in the previous layer. The mathematical model based on the transfer matrix method receives the structural parameters and amplitude coupling ratio data output by the one-dimensional convolution inverse design network. The amplitude coupling ratio prediction network outputs the amplitude coupling ratio data to calculate and generate a photon band image; the transmission spectrum forward prediction model includes a photon band encoder and a band-transmission spectrum decoder; the photon band encoder is used to receive the photon band image of size 267×200 calculated by the transfer matrix method, and encode it into a one-dimensional band image code with a length of 8192; the photon band encoder includes a convolutional layer and a fully connected layer; the band-transmission spectrum decoder is used to receive the one-dimensional band image code of 8192 output by the photon band encoder, and decode it back into one-dimensional transmission spectrum data with a length of 501; the band-transmission spectrum decoder includes a convolutional layer and a fully connected layer.

7. A topological optical switch structure design system according to claim 5, characterized in that: The mathematical model based on the transfer matrix method predicts the amplitude coupling ratio data of the network output according to the structural parameters and amplitude coupling ratio output by the one-dimensional convolution inverse design network, and uses the transfer matrix method and Bloch theorem to obtain the eigenvalue according to the eigenvalue equation, discards each eigenvalue that does not have a unit mode, and then scans the Bloch wave vector k x and phase, we get the photon band image, as shown in the formula: Among them, M x M is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the x direction, y is the transmission matrix representing the coupling relationship of the microring resonator coupling array in the y direction, k and τ are the amplitude coupling ratio and amplitude transmittance between the field ring and the connecting ring, respectively, and both satisfy |k| 2 +|τ| 2 =1, Φ is the phase shift produced when the optical path is one quarter of the micro-ring circumference; det(S) is The determinant of k x is the Bloch wave vector, b' is the input amplitude at the coupling point between the adjacent connecting ring and the field ring, and d' is the output amplitude at the coupling point between the adjacent connecting ring and the field ring.

8. A topological optical switch structure design system according to claim 5, characterized in that: The loss function is as follows: Among them, L δ (y,f(x)) is the loss function, y is the true parameter, f(x) is the predicted parameter, and δ is the parameter of Huber Loss.

Images