Optical system forward design method and model based on multi-branch weighted feature fusion

Abstract

The present application relates to a kind of optical system forward design method and model based on multi-branch weighted feature fusion, the method includes the following steps: the selection and construction of input feature;Multi-branch structure design;Self-adapting weighted feature fusion mechanism;The predicted output of forward design;Training and optimization process;The verification and application of output result.The optical system forward design method based on multi-branch weighted feature fusion network of the present application uses deep learning network to directly learn the mapping relationship between "performance index-design parameter", with multiple optical performance indexes (such as MTF curve, wavefront distribution diagram, Zernike aberration coefficient, etc.) as input, corresponding optical system design parameter is quickly output, truly realizes from optical system performance target to the forward prediction of structure parameter, avoids tedious, inefficient iterative optimization process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical system design and artificial intelligence, and particularly relates to an optical system forward design method and model based on multi-branch weighted feature fusion. BACKGROUND

[0002] As a core component in precision optoelectronic equipment, the performance of an optical system directly affects the imaging quality, detection accuracy and system stability. Traditional optical system design mainly relies on the experience accumulation and parameter tuning of optical engineers based on the theories of geometric optics and physical optics. The general design process includes: clearly defining system requirements, selecting a reasonable initial structure, using optical design software for parameter optimization, and then combining imaging quality evaluation and tolerance analysis to complete system finalization. However, in actual engineering applications, with the continuous development of optical systems towards high resolution, large field of view, wide waveband, miniaturization and other directions, the coupling relationship of internal parameters of the system is becoming increasingly complex, and the traditional design process based on human experience and numerical optimization faces the problems of low efficiency, large amount of calculation, long design cycle, and the design quality is significantly affected by the experience level of the designer.

[0003] In recent years, with the development of artificial intelligence, especially deep learning technology, new ideas have been provided for the intelligent design of optical systems. Researchers have begun to try to learn the mapping relationship between the design parameters and performance indicators of optical systems by constructing deep neural network models, to achieve intelligent prediction from performance targets to design parameters. This kind of method is usually based on a large number of historical design data, using performance indicators (such as wavefront aberration, MTF curve, etc.) to train network models, and combining optimization algorithms to deduce the optical design parameters that meet the performance requirements in the use stage. For example, the journal article "Hybrid IPSO-IAGA-BPNN algorithm-based rapid multi-objective optimization of a fully parameterized spaceborne primary mirror", published in Applied Optics, written by Tao Qin, Jun Li Guo, Zi Jian Jing, Pei Xian Han, and Bo Qi, proposes a hybrid algorithm based on improved particle swarm optimization, adaptive genetic algorithm and back propagation neural network. The model introduces adaptive inertia weight and improved genetic operators in particle swarm optimization and adaptive genetic algorithm, respectively, optimizes the connection parameters of the neural network using the improved algorithm to improve the global search ability of the algorithm, realizes the rapid multi-objective optimization design of the multi-aperture primary mirror, and establishes a fully parameterized primary mirror structure. The journal article "Direct generation of starting points for freeform off-axis three-mirror imaging system design using neural network based deep-learning", published in Optics Express, written by Tong Yang, Dewen Cheng, and Yongtian Wang, proposes a framework for generating initial points for freeform off-axis three-mirror imaging system design using neural network based deep-learning. The framework uses various system specifications and corresponding surface data obtained from system evolution as data sets to train the network. Generally, the network obtained can immediately generate good initial points for a specific system specification for further optimization. Although the above techniques have reduced the burden of manual work to some extent, there are still many limitations in this kind of reverse design method: first, it relies on subsequent optimization algorithms, resulting in a still cumbersome and time-consuming overall design process; second, error accumulation is easy to occur during the optimization process, affecting the design accuracy; third, the network structure mostly uses single path or shallow feature extraction, making it difficult to accurately model the complex relationship of multi-scale, multi-source, non-linear strong coupling in optical system design.

[0004] Therefore, how to construct a deep network architecture with strong expression ability and feature integration ability, break through the bottleneck of existing models in feature fusion and modeling precision, become the key technical problem of realizing efficient and reliable optical system intelligent design. SUMMARY

[0005] The present application aims to solve the technical problem of the bottleneck of the existing model for optical system design in feature fusion and modeling precision, and provides an optical system forward design method and model based on multi-branch weighted feature fusion.

[0006] In order to solve the above technical problems, the technical scheme of the present application is as follows:

[0007] An optical system forward design method based on multi-branch weighted feature fusion, comprising the following steps:

[0008] Step 1: selection and construction of input features;

[0009] The performance indicators of multiple optical systems are used as input features;

[0010] Step 2: multi-branch structure design;

[0011] A multi-branch deep learning network structure is adopted, and each branch of the network structure is responsible for extracting optical design features from different input feature dimensions or different scales;

[0012] Step 3: adaptive weighted feature fusion mechanism;

[0013] An adaptive weighted feature fusion mechanism is adopted, which automatically adjusts the weights of each input feature according to the contribution of different input features to the final design parameters, and realizes accurate prediction of the optical system design parameters;

[0014] Step 4: prediction output of forward design;

[0015] The forward design method is adopted to directly predict the corresponding optical system design parameters from multiple optical performance indicators;

[0016] Step 5: training and optimization process;

[0017] In the training process, the parameters of the multi-branch deep learning network structure model are optimized by minimizing the error function; the network weights are updated by the back propagation algorithm, and the high efficiency of the multi-branch deep learning network structure model in the design task is realized;

[0018] Step 6: verification and application of output results;

[0019] The trained multi-branch deep learning network structure model predicts optical system design parameters in real time by inputting different performance indicators; the output design parameters are used for actual manufacturing and debugging of the optical system, or provide reference for optimization strategy in the design iteration process.

[0020] In the above technical solution, the performance indicators of the plurality of optical systems in step 1 include Zernike aberration coefficients, wavefront distribution maps, modulation transfer functions, and point spread functions.

[0021] In the above technical solution, the large-scale optical design data trained in step 5 includes design parameters and corresponding performance indicators of different optical systems.

[0022] The network model applied to the above multi-branch weighted feature fusion-based optical system forward design method includes an input feature construction module, a multi-branch feature extraction module, an adaptive weighted feature fusion module, and an optical design parameter prediction module; wherein:

[0023] The input feature construction module is used to receive a plurality of performance indicators of an optical system and preprocess them into input features of the network model;

[0024] The multi-branch feature extraction module includes a Zernike branch, a wavefront map branch, and an MTF curve branch; wherein: the Zernike branch uses a multi-layer perceptron to process Zernike aberration coefficients to extract global design features therein; the wavefront map branch uses a multi-layer convolutional neural network to process wavefront distribution images to extract spatial distortion patterns thereof; the MTF curve branch uses a one-dimensional convolutional neural network or a frequency domain attention network to process frequency distribution features of the MTF curve ;

[0025] The adaptive weighted feature fusion module is used to dynamically fuse the features extracted by the Zernike branch, the wavefront map branch, and the MTF curve branch, considering the influence degree of different input sources on the final design;

[0026] The optical design parameter prediction module is used to receive the fused high-dimensional features and map them to the final optical design parameters through a multi-layer perceptron.

[0027] In the Zernike branch of the multi-branch feature extraction module in the above technical solution, the feature vector of the Zernike branch is given by the following formula:

[0028] ;

[0029] wherein, is the dimension of the feature space, represents a multi-layer perception, represents a space composed of real number vectors;

[0030] In the wavefront map branch of the multi-branch feature extraction module, the feature vector of the wavefront map branch obtained is represented as:

[0031] ;

[0032] wherein, represents a convolution operation, is the dimension of the feature space represents a space composed of real number vectors;

[0033] In the MTF curve branch of the multi-branch feature extraction module, the feature vector of the MTF curve branch obtained is represented as:

[0034] ;

[0035] wherein, represents the meaning, represents a convolution operation, is the dimension of the feature space, represents a space composed of real number vectors.

[0036] In the above technical solution, the adaptive weighted feature fusion module dynamically fuses the features, specifically:

[0037] Calculate the weight coefficient of each feature vector wherein, , the weight coefficient of each branch feature vector is calculated by the following formula:

[0038] ;

[0039] wherein, represents the learnable weight vector of the i-th branch, represents the feature vector of the i-th branch, extracted by the branch network, represents the branch index currently calculated, taking values , , , , , respectively represent the learnable weight vectors of the Zernike branch, the wavefront map branch and the MTF curve branch, , , respectively represent the feature vectors of the Zernike branch, the wavefront map branch and the MTF curve branch, represents the vector transposition operation, denotes the normalized attention weight coefficient;

[0040] After weighting, the fusion feature vector is given by:

[0041] ;

[0042] wherein, , , are normalized attention weight coefficients of the feature vectors , and respectively.

[0043] In the above technical solution, the prediction function of the optical design parameter prediction module is:

[0044] ;

[0045] wherein, is the predicted optical design parameter, is the mapping function of the multi-layer perception network.

[0046] In the above technical solution, the training process in the network model using process includes the following steps:

[0047] Step 1, collect optical system design samples to construct a data set, wherein, is the known design parameter;

[0048] Step 2, build the network architecture of the network model, take the performance index as the input, and the network model outputs the predicted parameter ;

[0049] Step 3, define the loss function for optimization:

[0050] ;

[0051] wherein, is the network model parameter, is the regularization term coefficient, denotes the sample index, denotes the total number of training samples;

[0052] Step 4, use the optimizer to perform gradient descent to update the network model parameter;

[0053] Step 5, evaluate the prediction accuracy and generalization ability of the network model on the validation set until convergence.

[0054] In the above technical solution, the model inference stage in the network model use process comprises the following steps:

[0055] Step 1, input the target optical system performance index combination;

[0056] Step 2, after being processed by the input feature construction module, enter the network model and pass through a forward propagation;

[0057] Step 3, the network model quickly outputs a set of optical design parameters ;

[0058] Step 4, the optical design parameters are directly used for preliminary optical system construction, or are used as initial solutions of optical design software for fine tuning.

[0059] The present application has the following beneficial effects:

[0060] The present application realizes true forward design of an optical system: specifically, the optical system forward design method based on multi-branch weighted feature fusion of the present application is different from the traditional method of relying on optimization algorithm for reverse design, adopts a deep learning network to directly learn the mapping relationship between performance indicators and design parameters, takes multiple optical performance indicators (such as MTF curve, wavefront distribution diagram, Zernike aberration coefficient, etc.) as input, and quickly outputs corresponding optical system design parameters, thereby truly realizing forward prediction from system performance target to structure parameter and avoiding a cumbersome and inefficient iterative optimization process.

[0061] The present application improves design efficiency and reduces design cost: the optical system forward design method based on multi-branch weighted feature fusion of the present application can directly output design parameters in the inference stage by constructing a deep neural network model with a parallel multi-branch structure, greatly reduces the trial-and-error parameter adjustment process and optimization calculation time in traditional design, significantly improves the optical system design efficiency, shortens the product development cycle, and thus reduces the overall research and development cost.

[0062] The present application enhances the expression ability and generalization ability of the model: the optical system forward design method based on multi-branch weighted feature fusion of the present application realizes effective extraction and adaptive fusion of design features from different scales and semantic levels by introducing a parallel multi-branch structure and a weighted feature fusion mechanism, enhances the modeling ability of the model for complex nonlinear design rules, and makes the model network structure show stronger generalization ability and robustness when facing different types of optical system design tasks.

[0063] The application improves prediction accuracy and stability: compared with a model relying on only a single performance indicator, the optical system forward design method based on multi-branch weighted feature fusion of the application fully integrates multi-dimensional performance information, so that the input features are more comprehensive and discriminative, which helps to improve the accuracy and stability of design parameter prediction, reduce design error rate, and improve the final optical system imaging quality.

[0064] The application reduces the dependence on artificial experience: the optical system forward design method based on multi-branch weighted feature fusion of the application adopts a data-driven modeling method, relies on large-scale design data to automatically learn parameter distribution and performance response relationship, and can complete high-quality forward design without deep optical design experience, which helps to promote the intelligentization, standardization and scaling development of optical design. BRIEF DESCRIPTION OF DRAWINGS

[0065] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0066] Figure 1 The model overall architecture diagram of the network model applied to the optical system forward design method based on multi-branch weighted feature fusion of the application.

[0067] Figure 2 The model training and inference process diagram of the network model applied to the optical system forward design method based on multi-branch weighted feature fusion of the application. DETAILED DESCRIPTION

[0068] The application idea of the application is:

[0069] The application proposes an optical system forward design method based on multi-branch weighted feature fusion, which realizes the key transformation of optical system design paradigm from "reverse solving" to "forward generation". The method takes multiple key performance indicators of the optical system as input features, including but not limited to Zernike aberration coefficients, wavefront distribution map, modulation transfer function, etc. Through constructing a deep neural network model with parallel multi-branch structure, design features are extracted from different scales and semantic levels, and a weighted feature fusion mechanism is adopted to adaptively integrate multi-branch features, which improves the modeling ability and feature expression ability of the model network structure to the design law. Finally, the model can realize fast prediction of the target optical system design parameters, and has the advantages of fast inference speed, high accuracy and strong stability. The application not only improves the efficiency and intelligent level of optical system forward design, but also significantly reduces the dependence on design experience, provides a new technical path for rapid development and large-scale iteration of high-performance optical systems, and has wide application prospect and engineering value.

[0070] The application will be described in detail below in combination with the drawings.

[0071] The optical system forward design method based on multi-branch weighted feature fusion of the application aims to fuse design features from different scales and semantic levels through a multi-branch deep neural network model, and directly realize forward prediction and optimization of optical system design parameters.

[0072] Step 1: selection and construction of input features;

[0073] The application takes multiple performance indicators of optical systems as input features, including but not limited to Zernike aberration coefficients, wavefront distribution maps, modulation transfer functions (MTF), point spread functions, etc. By comprehensively considering the key performance indicators of optical systems, the application ensures the comprehensive adaptability of the multi-branch deep learning network structure model to the design task. Compared with traditional methods, the application fully utilizes multi-dimensional and multi-source performance information to obtain more accurate optical system design parameters.

[0074] Step 2: multi-branch structure design;

[0075] The application adopts a multi-branch deep learning network structure, each branch of which is responsible for extracting optical design features from different input feature dimensions or different scales. This multi-branch structure can process data from multiple performance indicators in parallel and effectively extract features. The output features of each branch structure are then adaptively integrated through adaptive weighted feature fusion, optimizing the importance of different features and enhancing the expression ability of the multi-branch deep learning network structure model for complex nonlinear relationships.

[0076] Step 3: adaptive weighted feature fusion mechanism;

[0077] The adaptive weighted feature fusion mechanism automatically adjusts the weights of each input feature according to its contribution to the final design parameters, realizing accurate prediction of optical system design parameters.

[0078] Based on the multi-branch structure, the application designs an adaptive weighted feature fusion mechanism to optimize the synergistic effect between multi-level and different semantic features. The adaptive weighted feature fusion mechanism can automatically adjust the weights of each input feature according to its contribution to the final design parameters, thereby realizing accurate prediction of optical system design parameters. Through this mechanism, the multi-branch deep learning network structure model can better adapt to multi-element features and multiple targets in complex optical design tasks, further improving the stability and precision of the prediction results.

[0079] Step 4: prediction output of forward design;

[0080] Unlike traditional reverse design methods, the present application adopts a forward design approach to directly predict the corresponding optical system design parameters from multiple optical performance indicators. Through the trained multi-branch deep learning network structure model, a set of design parameters that meet the preset performance requirements can be directly outputted after inputting the performance indicators of the optical system, such as the shape of the optical element, the material selection, the relative position between the optical elements, etc.

[0081] Step 5: training and optimization process;

[0082] The present application adopts large-scale optical design data for training of the multi-branch deep learning network structure model, which includes design parameters and corresponding performance indicators of different optical systems. In the training process, the parameters of the multi-branch deep learning network structure model are optimized by minimizing the error function (such as mean square error or cross entropy), ensuring that the multi-branch deep learning network structure model can quickly converge in different design tasks and has strong generalization ability. Through the backpropagation algorithm, the network weights are updated to achieve high efficiency of the multi-branch deep learning network structure model in the design task.

[0083] Step 6: verification and application of output results;

[0084] The trained multi-branch deep learning network structure model can predict the optical system design parameters in real time by inputting different performance indicators (such as MTF, wavefront distribution map, etc.). The output design parameters can be further used for actual manufacturing and debugging of the optical system, or provide reference for optimization strategy in the design iteration process.

[0085] The present application provides an effective solution for intelligent design of optical systems.

[0086] In combination with Figure 1 Specific description: the network model applied to the optical system forward design method based on multi-branch weighted feature fusion of the present application is a deep neural network model with multi-branch structure, Figure 1 The overall architecture of the network model is shown, which is composed of the following four modules: input feature construction module, multi-branch feature extraction module, adaptive weighted feature fusion module and optical design parameter prediction module. The network model realizes efficient prediction of optical system design parameters by modeling features of multiple key performance indicators and fusing multi-source and multi-scale information.

[0087] The network model applied to the optical system forward design method based on multi-branch weighted feature fusion of the application comprises two stages of use process: training stage and inference stage. The network model structure, model training stage and model inference stage are introduced in detail below.

[0088] I. Network model structure composition.

[0089] a. Input feature construction module. This module is used to receive multiple performance indicators of the optical system and pre-process them into the input features of the model. The input features include but are not limited to the following important optical performance indicators:

[0090] 1) Zernike aberration coefficient: used to compress the representation of wavefront error, represented as a one-dimensional vector , wherein is the dimension of the Zernike coefficient, represents the Zernike aberration coefficient, represents a space composed of real vectors.

[0091] 2) Wavefront map: reflects the phase change of light wave in spatial domain, usually a two-dimensional image, represented as , wherein and represent the height and width of the wavefront map respectively, represents the wavefront map image, represents a space composed of real vectors.

[0092] 3) Modulation transfer function (MTF) curve: represents the spatial frequency response capability of the optical system, usually a one-dimensional vector , represented as , wherein is the number of data points of the MTF curve, represents a space composed of real vectors.

[0093] After receiving these performance indicators, the input features are pre-processed by normalization, completion or interpolation, and then sent to the corresponding branch structure network of the multi-branch feature extraction module for feature extraction.

[0094] b. Multi-branch feature extraction module. This module is composed of three structurally independent neural network branches, each branch corresponding to a different type of input performance indicator, Figure 1 wherein the Zernike branch is referred to as branch-Z, the wavefront map branch is referred to as branch-W, and the MTF curve branch is referred to as branch-M. Specifically:

[0095] 1) Branch-Z: Zernike coefficients are processed by a multi-layer perceptron (MLP) to extract global design features. The feature vector of Branch-Z is given by:

[0096]

[0097] where is the dimension of the feature space, denotes a multi-layer perceptron, and denotes a space of real-valued vectors.

[0098] 2) Branch-W: Wavefront images are processed by a multi-layer convolutional neural network (CNN) to extract spatial distortion patterns. The feature vector of Branch-W is given by:

[0099]

[0100] where denotes a convolution operation, is the dimension of the feature space, and denotes a space of real-valued vectors.

[0101] 3) Branch-M: MTF curves are processed by a one-dimensional convolutional neural network (1D-CNN) or a frequency-domain attention network to extract frequency distribution features. The feature vector of Branch-M is given by:

[0102]

[0103] where denotes a dimension, denotes a convolution operation, is the dimension of the feature space, and denotes a space of real-valued vectors.

[0104] The three branches extract deep semantic features of different performance indicators through their respective networks, outputting high-order feature vectors of the same dimension , and These vectors represent the depth information of the respective performance data.

[0105] ​​​​​​c. Adaptive weighted feature fusion module. This module dynamically fuses the features extracted by the three branches, considering the influence of different input sources on the final design. To this end, a weighted feature fusion method based on attention mechanism is adopted. Specifically, the weight coefficient of each feature vector is calculated wherein The weight coefficient of each branch feature vector is calculated by the following formula:

[0106] ;

[0107] wherein represents the learnable weight vector of the th branch, represents the feature vector of the th branch extracted by the branch network, represents the current calculated branch index, taking values , , , respectively represent the learnable weight vector of the Zernike branch, the wavefront map branch, and the MTF curve branch, , , respectively represent the feature vector of the Zernike branch, the wavefront map branch, and the MTF curve branch, represents the vector transposition operation, represents the normalized attention weight coefficient; this formula weights the features of each branch through a softmax operation to reflect the importance of each input in the final design. After weighting, the fusion feature vector is given by the following formula:

[0108] ;

[0109] wherein , , respectively are the normalized attention weight coefficients of the feature vectors , and ;

[0110] This mechanism can dynamically adjust the importance of each performance feature according to different task scenarios, thereby improving the generalization ability and accuracy of the network model.

[0111] d. Optical design parameter prediction module. This module is the output module of the network model, which receives the high-dimensional fusion feature vector after fusion, and maps it to the final optical design parameter through a multi-layer perceptron (MLP). Specifically, the prediction function is:

[0112] ;

[0113] wherein, is the predicted optical design parameter, is the mapping function of the multi-layer perceptron network. The output optical design parameters include but are not limited to: the radius of curvature of each lens, the thickness of the lens, the refractive index of the lens, the stop position, the total length of the optical system, the field of view angle and other key structural parameters. Through the prediction of this module, the network model can directly generate design parameters that meet the given optical performance targets, avoiding the complex inverse optimization process in traditional methods.

[0114] II. Model training phase.

[0115] The goal of the training phase is to perform end-to-end supervised learning of the network model based on a large number of existing optical design sample data. The training process is shown in Figure 2 (Figure only shows the step outline), including the following steps:

[0116] 1) Collect optical system design samples and build a dataset (i.e. training set), wherein, is the known design parameter;

[0117] 2) Build the network architecture of the network model, taking the performance indicators as input, and the network model outputs the predicted parameters ;

[0118] 3) Define the loss function (such as mean square error) for optimization:

[0119] ;

[0120] wherein, is the network model parameter, is the regularization term coefficient, represents the sample index, represents the total number of training samples;

[0121] 4) Select an optimizer (such as Adam) to perform gradient descent and update the network model parameters;

[0122] 5) Evaluate the prediction accuracy and generalization ability of the network model on the validation set until convergence.

[0123] III. Model inference phase.

[0124] The model inference phase is shown in Figure 2 (Figure only shows the step outline), including the following steps:

[0125] 1) Input the target optical system performance indicator combination;

[0126] 2) Load network model parameters, and the performance index processed by the input feature construction module enters the network model and is forward propagated once;

[0127] 3) The network model quickly outputs a set of optical design parameters ;

[0128] 4) The optical design parameters can be directly used for preliminary optical system construction, or used as the initial solution of traditional optical design software for fine tuning.

[0129] Figure 2 The process of model training and model inference of the network model is shown. First, the model training is completed by collecting samples, constructing a data set, building a network, defining a loss function, and optimizing parameters; then, the target performance index is input, the model is loaded and forward propagated, and the design parameters are obtained, thereby realizing the automatic construction of the optical system.

[0130] In summary, the optical system forward design method and model based on multi-branch weighted feature fusion of the present application can not only improve the efficiency and intelligent level of optical system forward design, but also significantly reduce the dependence on design experience, providing a new technical path for the rapid development and large-scale iteration of high-performance optical systems, and has wide application prospects and engineering value.

[0131] Obviously, the above embodiments are only examples for clear illustration, and not limitation on the embodiments. For ordinary skilled in the art, other different forms of changes or variations can be made on the basis of the above description. Here, all the embodiments need not and cannot be exhausted. The obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A method for optical system forward design based on multi-branch weighted feature fusion, characterized in that, The method comprises the following steps: Step 1: selection and construction of input features; Taking performance indicators of multiple optical systems as input features; Step 2: multi-branch structure design; A multi-branch deep learning network structure is adopted, and each branch of the network structure is responsible for extracting optical design features from different input feature dimensions or different scales; Step 3: adaptive weighted feature fusion mechanism; An adaptive weighted feature fusion mechanism is adopted to automatically adjust the weights of each input feature according to the contribution of different input features to the final design parameters; Step 4: prediction output of forward design; A forward design method is adopted to directly predict the optical system design parameters from multiple optical performance indicators; Step 5: training and optimization process; In the training process, the parameters of the multi-branch deep learning network structure model are optimized by minimizing the error function; the network weights are updated by the back propagation algorithm; Step 6: verification and application of output results; The trained multi-branch deep learning network structure model can predict optical system design parameters in real time by inputting different performance indicators; The output design parameters are used for the actual manufacturing and debugging of optical systems, or provide reference for optimization strategy in the design iteration process; The network model applied to the optical system forward design method based on multi-branch weighted feature fusion comprises an input feature construction module, a multi-branch feature extraction module, an adaptive weighted feature fusion module, and an optical design parameter prediction module; wherein: The input feature construction module is used to receive multiple performance indicators of an optical system and preprocess them as input features of the network model; The multi-branch feature extraction module comprises a Zernike branch, a wavefront map branch and an MTF curve branch; wherein: the Zernike branch adopts a multi-layer perception to process Zernike aberration coefficients The global design features are extracted; the wavefront map branch uses a multi-layer convolutional neural network to process a wavefront distribution image , and extracts the spatial distortion mode thereof; the MTF curve branch adopts a one-dimensional convolutional neural network or a frequency domain attention network to process the frequency distribution characteristics of the MTF curve ; The adaptive weighted feature fusion module is used to dynamically fuse the features extracted by the Zernike branch, the wavefront graph branch and the MTF curve branch; The optical design parameter prediction module is used to receive the fused high-dimensional features and map them to the final optical design parameters through a multi-layer perception machine In the Zernike branch of the multi-branch feature extraction module, the feature vector of the Zernike branch is given by the equation: ； wherein, is the dimension of the feature space, denotes a multi-layer perceptron, denotes a space of real-valued vectors; In the wave front map branch of the multi-branch feature extraction module, the feature vector of the wave front map branch obtained is represented as: ； wherein, denotes a convolution operation, is the dimension of the feature space, denotes a space of real-valued vectors; In the branch of the MTF curve of the multi-branch feature extraction module, the feature vector of the obtained branch of the MTF curve is represented as: ； wherein, denotes a dimension, denotes a convolution operation, is a dimension of a feature space, denotes a space of real vectors.

2. The optical system forward design method based on multi-branch weighted feature fusion according to claim 1, characterized in that, The multiple performance indicators of the optical system in step 1 include Zernike aberration coefficients, wavefront distribution graphs, modulation transfer functions, and point spread functions.

3. The optical system forward design method based on multi-branch weighted feature fusion according to claim 1, characterized in that, The large-scale optical design data used for training in step 5 includes design parameters and corresponding performance indicators of different optical systems.

4. The optical system forward design method based on multi-branch weighted feature fusion according to claim 1, characterized in that, The adaptive weighted feature fusion module dynamically fuses features, specifically: calculating a weight coefficient of each feature vector wherein, calculating a weight coefficient of each branch feature vector by the following equation: ； wherein, represents the learnable weight vector of the th branch, represents the feature vector of the th branch, represents the branch index of the current calculation, taking values , , , respectively represent the learnable weight vector of the Zernike branch, the wavefront map branch, the MTF curve branch, , , respectively represent the feature vector of the Zernike branch, the wavefront map branch, the MTF curve branch, represents the vector transposition operation, represents the normalized attention weight coefficient; After weighting, the fusion feature vector is given by: ； wherein, , , are normalized attention weight coefficients of the feature vectors , and respectively.

5. The optical system forward design method based on multi-branch weighted feature fusion according to claim 4, characterized in that, The prediction function of the optical design parameter prediction module is: ； wherein, is a predicted optical design parameter, is a mapping function of a multi-layer perceptron network.

6. The optical system forward design method based on multi-branch weighted feature fusion according to claim 1, characterized in that, The training process of the network model uses the following steps: Step 1, collect optical system design samples, build dataset, wherein, are known design parameters; Step 2, build the network architecture of the network model, take the performance index as input, and the network model outputs the prediction parameter ; Step 3, define loss function Perform optimization: ； wherein, are network model parameters, are regularization coefficients, denotes a sample index, denotes the total number of training samples; Step 4: use the optimizer to perform gradient descent and update the network model parameters; Step 5: evaluate the prediction accuracy and generalization ability of the network model on the validation set until convergence.

7. The optical system forward design method based on multi-branch weighted feature fusion according to claim 1, characterized in that, The model inference stage of the network model uses the following steps: Step 1: input the target optical system performance indicator combination; Step 2: after being processed by the input feature construction module, enter the network model and perform a forward propagation. Step 3, the network model quickly outputs a set of optical design parameters ; Step 4, the optical design parameters are directly used for the primary optical system construction, or fine-tuned as the initial solution of the optical design software. directly used for the primary optical system construction, or fine-tuned as the initial solution of the optical design software.

Images