CN-121981052-A - Polarizer design method and device based on dual-path collaborative training

CN121981052ACN 121981052 ACN121981052 ACN 121981052ACN-121981052-A

Abstract

The application relates to a design method and a device of a polarizer based on dual-path collaborative training, which are characterized in that the combination characteristics between the structural parameters and the quality factors of the polarizer and the combination characteristics between the two-dimensional material parameters and the quality factors of the polarizer are mapped to a shared potential space together by introducing an encoder and a priori network, a forward decoder and a reverse decoder work cooperatively in the potential space, and knowledge fusion of forward prediction from parameters to performances and reverse prediction from performances to parameters is realized by sampling potential vectors, so that the performances of known structures can be rapidly predicted, and waveguide parameters can be obtained according to expected performances and two-dimensional material parameters. The application realizes knowledge fusion of forward and reverse design in shared potential space through a dual-path cooperative training mechanism, thereby effectively solving the problems of large simulation resource consumption and redundancy of design frames in the traditional polarizer design and providing a way for efficient and generalized design of integrated waveguide polarizers.

Inventors

WANG RONG
HUANG DUAN
ZHANG HANG
ZHANG LING
XIAO HONGTAO
ZHANG YUNJIE
GUO JIAWEI

Assignees

中南大学

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. A method of designing a polarizer based on dual-path co-training, the method comprising: The method comprises the steps of obtaining a prediction model and a first training set, wherein the prediction model comprises an encoder, a priori network, a forward decoder and a reverse decoder, and the first training set comprises structural parameters and quality factors of a polarizer; Training the predictive model according to the first training set, wherein the training process comprises the steps of extracting a first feature between a structural parameter and a quality factor based on the encoder and mapping the first feature to a potential space, extracting a second feature between a two-dimensional material parameter and the quality factor based on the prior network and mapping the second feature to the potential space, sampling a first potential vector from the potential space according to posterior distribution of the encoder and outputting a predicted quality factor based on the first potential vector and training parameters by the forward decoder, sampling a second potential vector from the potential space according to prior distribution of the prior network and outputting a predicted waveguide parameter based on the second potential vector and the two-dimensional material parameter by the reverse decoder; in response to a parameter prediction instruction of a target polarizer, in the case that a predicted parameter is a figure of merit, a structural parameter of the target polarizer is input into the prediction model to sample a first target potential vector from the potential space based on a gaussian distribution, a figure of merit of the target polarizer is output according to the first target potential vector based on the forward decoder, in the case that a predicted parameter is a waveguide parameter, a two-dimensional material parameter and a figure of merit of the target polarizer are input into the prediction model to sample a second target potential vector from the potential space based on an a priori distribution of the a priori network, and a waveguide parameter of the target polarizer is output according to the second target potential vector based on the reverse decoder.
2. A method of designing a polarizer based on dual-path co-training as defined in claim 1, wherein the number of first target potential vectors is plural; the outputting, based on the forward decoder, a figure of merit for the target polarizer according to the first target potential vector, comprising: combining the structural parameters of the target polarizer with each of the first target potential vectors, respectively; Predicting a corresponding figure of merit based on the forward decoder from each combination; The average of all the predicted figures of merit is calculated as the figure of merit for the target polarizer.
3. The method of designing a polarizer based on dual-path co-training of claim 1, wherein the number of second target potential vectors is plural; The outputting waveguide parameters of the target polarizer according to the second target potential vector based on the inverse decoder comprises: combining a two-dimensional material parameter and a figure of merit with each of the second target potential vectors, respectively; based on the inverse decoder, a plurality of waveguide parameters of the target polarizer are predicted from each combination.
4. A method of designing a polarizer based on dual path co-training according to claim 1, wherein the predictive model loss function during training comprises: a KL divergence constraint that measures similarity between a posterior distribution of the encoder and an a priori distribution of the a priori network; calculating an error constraint of an error between the predicted figure of merit and the true figure of merit; and calculating an error constraint for an error between the predicted waveguide parameter and the actual waveguide parameter.
5. A method of designing a dual path co-training based polarizer according to claim 4, wherein the extracting a first feature between a structural parameter and a figure of merit based on the encoder and mapping the first feature to a potential space comprises: vectorizing the structural parameters to obtain structural parameter vectors; Normalizing the structural parameter vector; calculating posterior distribution mapped to potential space based on the normalized structural parameter vector and the quality factor by a full connection layer, wherein the posterior distribution comprises a mean value and a logarithmic variance; The extracting a second feature between a two-dimensional material parameter and a figure of merit based on the prior network, mapping the second feature to the potential space, comprising: Vectorizing the two-dimensional material parameters to obtain a material parameter vector; Normalizing the material parameter vector; and calculating prior distribution mapped to potential space based on the full connection layer according to the normalized material parameter vector and the quality factor, wherein the prior distribution comprises a mean value and a logarithmic variance.
6. A method of designing a polarizer based on dual path co-training according to claim 1, wherein prior to the parameter prediction instruction in response to a target polarizer, the method further comprises: under the condition that the prediction model training is completed and the polarizer of the first training set is made of a first two-dimensional material, acquiring a second training set composed of structural parameters and quality factors of the polarizer made of a second two-dimensional material; and training the prediction model according to the second training set based on the trained pre-training model parameters of the prediction model as initial values so as to obtain a trained new prediction model.
7. The method of designing a dual-path co-training based polarizer of claim 6, wherein the training the predictive model according to the second training set comprises: a KL divergence constraint that measures similarity between a posterior distribution of the encoder and an a priori distribution of the a priori network; calculating an error constraint of an error between the predicted figure of merit and the true figure of merit; calculating an error constraint of an error between the predicted waveguide parameter and the real waveguide parameter; And calculating regularization constraints of the degree of deviation between the current model parameters of the predictive model and the pre-trained model parameters.
8. A dual-path co-training based polarizer design apparatus, the apparatus comprising: The system comprises a data acquisition module, a prediction model and a first training set, wherein the prediction model comprises an encoder, a priori network, a forward decoder and a reverse decoder, the first training set comprises structural parameters and quality factors of a polarizer, and the structural parameters comprise two-dimensional material parameters and waveguide parameters; The model training module is used for training the prediction model according to the first training set, and the training process comprises the steps of extracting first features between structural parameters and quality factors based on the encoder and mapping the first features to potential space, extracting second features between two-dimensional material parameters and the quality factors based on the prior network, mapping the second features to the potential space, sampling first potential vectors from the potential space according to posterior distribution of the encoder and outputting predicted quality factors based on the first potential vectors and training parameters by the forward decoder, sampling second potential vectors from the potential space according to prior distribution of the prior network and outputting predicted waveguide parameters based on the reverse decoder according to the second potential vectors and the two-dimensional material parameters; The device comprises a parameter prediction module, a prediction model and a reverse decoder, wherein the parameter prediction module is used for responding to a parameter prediction instruction of a target polarizer, inputting structural parameters of the target polarizer into the prediction model to sample a first target potential vector from the potential space based on Gaussian distribution under the condition that the predicted parameters are quality factors, outputting the quality factors of the target polarizer according to the first target potential vector based on the forward decoder, inputting two-dimensional material parameters and the quality factors of the target polarizer into the prediction model to sample a second target potential vector from the potential space based on the prior distribution of the prior network under the condition that the predicted parameters are waveguide parameters, and outputting the waveguide parameters of the target polarizer according to the second target potential vector based on the reverse decoder.
9. An electronic device comprising at least one control processor and a memory communicatively coupled to the at least one control processor, the memory storing instructions executable by the at least one control processor to enable the at least one control processor to perform a dual path co-training based polarizer design method of any one of claims 1-7.
10. A computer-readable storage medium having stored thereon computer-executable instructions for causing a computer to perform a dual-path co-training based polarizer design method as recited in any one of claims 1to 7.

Description

Polarizer design method and device based on dual-path collaborative training Technical Field The embodiment of the application relates to the technical field of waveguide polarizer design, in particular to a polarizer design method and device based on dual-path collaborative training. Background The integrated waveguide polarizer is used as a core passive device in the photoelectric fields of optical communication, optical sensing and the like, the performance of the integrated waveguide polarizer directly determines the precision of optical signal transmission and regulation, and the integrated waveguide polarizer is a key support for the development of miniaturization and high performance of integrated photon chips. The core design goal of the device is to realize the efficient distinction between Transverse Electric (TE) and Transverse Magnetic (TM) polarization modes, the quality Factor (FOM) is a core index for measuring the performance of the device, and the performance of the device is determined by two-dimensional material optical parameters (thickness, real part/imaginary part of refractive index) and central waveguide dimension parameters (height and width). At present, integrated waveguide polarizers are mostly prepared by adopting silicon, silicon dioxide, silicon nitride and the like as waveguide and substrate materials and matching with two-dimensional materials such as graphene, graphene oxide, transition metal chalcogenide and the like, and the design core is to realize accurate matching of the materials and waveguide parameters so as to obtain optimal polarization performance. In the prior art, the design of an integrated waveguide polarizer mainly depends on commercial simulation software, and device design and performance evaluation are realized through full-parameter space scanning simulation. However, the existing solutions still have a plurality of technical defects, which are difficult to meet the high-efficiency and generalized design requirements of the photoelectric device, firstly, the simulation resources are high in consumption, a large amount of calculation force is required for full-parameter space scanning and repeated trial-and-error simulation, the design period is long, especially, the multiple adjustment and verification of parameters in reverse design further increase the time and calculation cost, secondly, the design framework is redundant, the forward (parameter to performance) and reverse (performance to parameter) designs adopt independent models or processes, the inherent association of the bidirectional mapping of the parameters and the performance is ignored, the knowledge sharing cannot be realized, and the overall design efficiency is limited. Disclosure of Invention The application provides a polarizer design method based on dual-path co-training, which comprises the following steps: The method comprises the steps of obtaining a prediction model and a first training set, wherein the prediction model comprises an encoder, a priori network, a forward decoder and a reverse decoder, and the first training set comprises structural parameters and quality factors of a polarizer; Training the predictive model according to the first training set, wherein the training process comprises the steps of extracting a first feature between a structural parameter and a quality factor based on the encoder and mapping the first feature to a potential space, extracting a second feature between a two-dimensional material parameter and the quality factor based on the prior network and mapping the second feature to the potential space, sampling a first potential vector from the potential space according to posterior distribution of the encoder and outputting a predicted quality factor based on the first potential vector and training parameters by the forward decoder, sampling a second potential vector from the potential space according to prior distribution of the prior network and outputting a predicted waveguide parameter based on the second potential vector and the two-dimensional material parameter by the reverse decoder; in response to a parameter prediction instruction of a target polarizer, in the case that a predicted parameter is a figure of merit, a structural parameter of the target polarizer is input into the prediction model to sample a first target potential vector from the potential space based on a gaussian distribution, a figure of merit of the target polarizer is output according to the first target potential vector based on the forward decoder, in the case that a predicted parameter is a waveguide parameter, a two-dimensional material parameter and a figure of merit of the target polarizer are input into the prediction model to sample a second target potential vector from the potential space based on an a priori distribution of the a priori network, and a waveguide parameter of the target polarizer is output according to the second target potential vector based on