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

Abstract

The invention relates to an optical system forward design method and model based on multi-branch weighted feature fusion. The method comprises the following steps: selecting and constructing input features; designing a multi-branch structure; a self-adaptive weighted feature fusion mechanism; predicting and outputting the forward design; a training and optimizing process; and verifying and applying an output result. According to the optical system forward design method based on the multi-branch weighted feature fusion network, a deep learning network is adopted to directly learn a mapping relation between performance indexes and design parameters, various optical performance indexes (such as an MTF curve, a wavefront distribution diagram and a Zernike aberration coefficient) are used as input, corresponding optical system design parameters are rapidly output, and the optical system forward design method based on the multi-branch weighted feature fusion network is achieved. Forward prediction from the performance target of the optical system to the structural parameters is truly realized, and a tedious and low-efficiency iterative optimization process is avoided.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology of optical system design and artificial intelligence, and in particular to a forward design method and model of an optical system based on multi-branch weighted feature fusion. Background Art

[0002] As the core component of precision optoelectronic equipment, the performance of the 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 theory of geometric optics and physical optics. The usual design process includes: clarifying the system requirements, selecting a reasonable initial structure, using optical design software to optimize parameters, and then combining imaging quality evaluation and tolerance analysis to complete the system finalization. However, in actual engineering applications, as optical systems continue to develop in the direction of high resolution, large field of view, wide band, miniaturization, etc., the coupling relationship between the internal parameters of the system is becoming increasingly complex. The traditional design process based on manual experience and numerical optimization faces the problems of low efficiency, large computational complexity, and long design cycle, and the design quality is significantly affected by the designer's level of experience.

[0003] In recent years, the development of artificial intelligence, particularly deep learning, has provided new insights into the intelligent design of optical systems. Researchers have begun exploring the mapping relationship between optical system design parameters and performance indicators by constructing deep neural network models, thereby achieving intelligent prediction from performance targets to design parameters. These methods typically rely on a large amount of historical design data, using performance indicators (such as wavefront aberrations and MTF curves) to train the network model. During the operational phase, optimization algorithms are then used to reversely derive optical design parameters that meet performance requirements. For example, the journal article "Hybrid IPSO-IAGA-BPNN algorithm-based rapid multi-objective optimization of a fully parameterized spaceborne primary mirror" (Applied Optics) by Tao Qin, JunLi Guo, ZiJianJing, PeiXian Han, and Bo Qi proposes a proxy model based on a hybrid algorithm of improved particle swarm optimization, an adaptive genetic algorithm, and an optimized back-propagation neural network. This model introduces adaptive inertia weights and improved genetic operators into particle swarm optimization and adaptive genetic algorithms, respectively. The improved algorithm is used to optimize the connection parameters of the neural network to improve the global search capability of the algorithm, achieving rapid multi-objective optimization design of a multi-aperture primary mirror and establishing 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" from Optics Express, by Tong Yang, Dewen Cheng, and Yongtian Wang, proposes a framework for generating starting 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 obtained network can immediately generate good starting points for specific system specifications for further optimization. Although the above-mentioned technology has reduced the manual burden to a certain extent, this type of reverse design method also has many limitations: first, it relies on subsequent optimization algorithms, resulting in the overall design process being cumbersome and time-consuming; second, errors are easily accumulated during the optimization process, affecting the design accuracy; third, the network structure mostly adopts a single path or shallow feature extraction, which makes it difficult to accurately model the complex relationships of multi-scale, multi-source, and nonlinear strong coupling in optical system design.

[0004] Therefore, how to build a deep network architecture with strong expression and feature integration capabilities and break through the bottleneck of existing models in feature fusion and modeling accuracy has become a key technical issue in achieving efficient and reliable intelligent design of optical systems. Summary of the Invention

[0005] The present invention aims to solve the technical problem that the models used for optical system design in the prior art have bottlenecks in feature fusion and modeling accuracy, 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 solutions of the present invention are as follows:

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

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

[0009] Taking the performance indicators of multiple optical systems as input features;

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

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

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

[0013] Adopting an adaptive weighted feature fusion mechanism, the weight of each input feature is automatically adjusted according to its contribution to the final design parameters, thus achieving accurate prediction of optical system design parameters.

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

[0015] Using forward design, the corresponding optical system design parameters are directly predicted from multiple optical performance indicators;

[0016] Step 5: Training and optimization process;

[0017] During 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 through the back-propagation algorithm to achieve efficient performance of the multi-branch deep learning network structure model in the designed task;

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

[0019] The trained multi-branch deep learning network structure model predicts the 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 the optical system, or provide a reference for optimization strategies during the design iteration process.

[0020] In the above technical solution, the performance indicators of the multiple optical systems in step 1 include: Zernike aberration coefficient, wavefront distribution diagram, modulation transfer function, and point spread function.

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

[0022] The network model applied to the above-mentioned optical system forward design method based on multi-branch weighted feature fusion 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 multiple performance indicators of the optical system and preprocess them into input features of the network model;

[0024] The multi-branch feature extraction module includes: Zernike branch, wavefront branch and MTF curve branch; among them: Zernike branch uses multi-layer perceptron to extract Zernike aberration coefficient Processing is performed to extract the global design features; the wavefront image branch uses a multi-layer convolutional neural network to process the wavefront distribution image , extract its spatial distortion pattern; the MTF curve branch uses a one-dimensional convolutional neural network or a frequency domain attention network to process the MTF curve Frequency distribution characteristics of

[0025] The adaptive weighted feature fusion module is used to dynamically fuse the features extracted from the Zernike branch, wavefront branch, and MTF curve branch, taking into account the impact 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 into the final optical design parameters through a multi-layer perceptron.

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

[0028] ;

[0029] in, is the dimension of the feature space, represents a multilayer perceptron, represents the space of real vectors;

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

[0031] ;

[0032] in, represents the convolution operation, is the dimension of the feature space represents the space of real vectors;

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

[0034] ;

[0035] in, Indicates meaning, represents the convolution operation, is the dimension of the feature space, Represents the space of real vectors.

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

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

[0038] ;

[0039] in, Indicates the The learnable weight vectors of the branches, Indicates the The feature vector of each branch is extracted by the branch network. Indicates the branch index of the current calculation, the value is , 、 、 Represent the learnable weight vectors of the Zernike branch, wavefront branch, and MTF curve branch, respectively. , , Represent the characteristic vectors of Zernike branch, wavefront branch, and MTF curve branch respectively, represents the vector transpose operation, Represents the normalized attention weight coefficient;

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

[0041] ;

[0042] in, , , The eigenvectors are 、 and The normalized attention weight coefficient.

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

[0044] ;

[0045] in, are the predicted optical design parameters, is the mapping function of the multilayer perceptron network.

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

[0047] Step 1: Collect optical system design samples and construct a data set, where: are known design parameters;

[0048] Step 2: Build the network architecture of the network model, take the performance indicators as input, and the network model outputs the prediction parameters ;

[0049] Step 3: Define the loss function To optimize:

[0050] ;

[0051] in, are network model parameters, is the regularization coefficient, represents the sample index, Represents the total number of training samples;

[0052] Step 4: Use the optimizer to perform gradient descent and update the network model parameters;

[0053] Step 5: The network model evaluates prediction accuracy and generalization ability on the validation set until convergence.

[0054] In the above technical solution, the model inference phase in the network model usage process includes 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, it enters the network model and undergoes one forward propagation;

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

[0058] Step 4: Optical design parameters It can be used directly for preliminary optical system construction or as an initial solution for fine-tuning in optical design software.

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

[0060] The present invention realizes forward design of optical systems in a true sense: specifically, the forward design method of optical systems based on multi-branch weighted feature fusion of the present invention is different from the traditional method of reverse design that relies on optimization algorithms. It adopts a deep learning network to directly learn the mapping relationship between "performance indicators-design parameters", and uses multiple optical performance indicators (such as MTF curves, wavefront distribution diagrams, Zernike aberration coefficients, etc.) as input to quickly output corresponding optical system design parameters, truly realizing forward prediction from system performance targets to structural parameters, and avoiding tedious and inefficient iterative optimization processes.

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

[0062] The present invention enhances the expressive power and generalization ability of the model: The optical system forward design method based on multi-branch weighted feature fusion of the present invention realizes the 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, thereby enhancing the model's ability to model complex nonlinear design rules, and enabling the model network structure to exhibit stronger generalization ability and robustness when facing different types of optical system design tasks.

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

[0064] The present invention reduces dependence on manual experience: the optical system forward design method based on multi-branch weighted feature fusion of the present invention adopts a data-driven modeling approach and relies on large-scale design data to automatically learn the relationship between parameter distribution and performance response. It can complete high-quality forward design without the need for profound optical design experience, which helps to promote the intelligent, standardized and large-scale development of optical design. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

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

[0067] Figure 2 The figure is a schematic diagram of the model training and reasoning process of the network model applied to the optical system forward design method based on multi-branch weighted feature fusion of the present invention. DETAILED DESCRIPTION

[0068] The inventive concept of the present invention is:

[0069] The present invention proposes a forward design method for optical systems based on multi-branch weighted feature fusion, which realizes the key transformation of the optical system design paradigm from "reverse solution" to "forward generation". This method uses multiple key performance indicators of the optical system as input features, including but not limited to Zernike aberration coefficients, wavefront distribution diagrams, modulation transfer functions, etc., and extracts design features from different scales and semantic levels by constructing a deep neural network model with a parallel multi-branch structure, and uses a weighted feature fusion mechanism to adaptively integrate multi-branch features, thereby improving the modeling ability and feature expression ability of the model network structure for design rules. Ultimately, the model can achieve rapid prediction of the design parameters of the target optical system, and has the advantages of fast reasoning speed, high accuracy and strong stability. The present invention can not only improve the efficiency and intelligence level of the forward design of the optical system, 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 broad application prospects and engineering value.

[0070] The present invention will be described in detail below with reference to the accompanying drawings.

[0071] The optical system forward design method based on multi-branch weighted feature fusion of the present invention aims to directly achieve forward prediction and optimization of optical system design parameters by fusing design features from different scales and semantic levels through a multi-branch deep neural network model. The optical system forward design method comprises the following steps:

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

[0073] This method uses multiple optical system performance indicators as input features, including but not limited to Zernike aberration coefficients, wavefront profiles, modulation transfer functions (MTFs), and point spread functions. By comprehensively considering key optical system performance indicators, the method ensures the comprehensive adaptability of the multi-branch deep learning network model to design tasks. Compared with traditional methods, this method 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 present invention utilizes a multi-branch deep learning network structure, in which each branch is responsible for extracting optical design features from different input feature dimensions or scales. This multi-branch structure enables parallel processing of data from multiple performance indicators and efficient feature extraction. The output features of each branch are then adaptively integrated through adaptive weighted feature fusion, optimizing the importance of different features and enhancing the multi-branch deep learning network model's ability to express complex nonlinear relationships.

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

[0077] Adopting an adaptive weighted feature fusion mechanism, the weight of each input feature is automatically adjusted according to its contribution to the final design parameters, thus achieving accurate prediction of optical system design parameters.

[0078] Based on a multi-branch architecture, the present invention designs an adaptive weighted feature fusion mechanism to optimize the synergy between multi-level, semantically distinct features. This adaptive weighted feature fusion mechanism automatically adjusts the weights of each input feature based on its contribution to the final design parameter, thereby enabling accurate prediction of optical system design parameters. This mechanism enables a multi-branch deep learning network model to better adapt to the diverse features and multiple objectives encountered in complex optical design tasks, further improving the stability and accuracy of prediction results.

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

[0080] Unlike traditional reverse design methods, this invention uses a forward design approach to directly predict the corresponding optical system design parameters from multiple optical performance indicators. The trained multi-branch deep learning network model, when fed with the optical system's performance indicators, directly outputs a set of design parameters that meet the preset performance requirements, such as the shape, material selection, and relative positioning of optical components.

[0081] Step 5: Training and optimization process;

[0082] This method trains a multi-branch deep learning network model using large-scale optical design data. This data includes design parameters and corresponding performance metrics for different optical systems. During training, the parameters of the multi-branch deep learning network model are optimized by minimizing an error function (such as mean squared error or cross-entropy). This ensures that the multi-branch deep learning network model converges quickly across various design tasks and exhibits strong generalization capabilities. A backpropagation algorithm is used to update network weights, achieving efficient performance across design tasks.

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

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

[0085] The optical system forward design method based on multi-branch weighted feature fusion of the present invention can achieve efficient and accurate optical system forward design without relying on complex optimization processes through a multi-branch weighted feature fusion network structure, significantly improve design efficiency and system performance, reduce design costs, and provide an effective solution for the intelligent design of optical systems.

[0086] Combine Figure 1 Specific description: The network model of the optical system forward design method based on multi-branch weighted feature fusion applied to the present invention is a deep neural network model with a multi-branch structure. Figure 1 The overall architecture of the network model is presented, consisting of four modules: an input feature construction module, a multi-branch feature extraction module, an adaptive weighted feature fusion module, and an optical design parameter prediction module. The network model efficiently predicts optical system design parameters by modeling features for multiple key performance indicators and fusing multi-source and multi-scale information.

[0087] The network model used in the optical system forward design method based on multi-branch weighted feature fusion in this invention comprises a two-stage workflow: a training phase and an inference phase. The following describes in detail the network model structure, the model training phase, and the model inference phase.

[0088] 1. Network model structure composition.

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

[0090] 1) Zernike aberration coefficient: used to compress and characterize the wavefront error, expressed as a one-dimensional vector ,in, is the dimension of the Zernike coefficients, represents the Zernike aberration coefficient, Represents the space of real vectors.

[0091] 2) Wavefront Map: reflects the phase change of light waves in the spatial domain, usually a two-dimensional image, represented by ,in, and Represent the height and width of the wavefront distribution diagram, represents the wavefront distribution image, Represents the space of real vectors.

[0092] 3) Modulation Transfer Function (MTF) curve: Characterizes the spatial frequency response capability of an optical system, usually a one-dimensional vector , expressed as ,in, is the number of data points of the MTF curve, Represents the space of real vectors.

[0093] After receiving these performance indicators, the input features are preprocessed 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 consists of three independent neural network branches, each corresponding to different types of input performance indicators. Figure 1 In this paper, the Zernike branch is referred to as branch-Z, the wavefront branch is referred to as branch-W, and the MTF curve branch is referred to as branch-M. Specifically:

[0095] 1) Zernike branch (Branch-Z): Multi-layer perceptron (MLP) is used to calculate the Zernike aberration coefficients. Process and extract the global design features. The eigenvector of the Zernike branch It is given by the following formula:

[0096] ;

[0097] in, is the dimension of the feature space, represents a multilayer perceptron, Represents the space of real vectors.

[0098] 2) Wavefront image branch (Branch-W): Use a multi-layer convolutional neural network (CNN) to process the wavefront image , extract its spatial distortion pattern. The eigenvector of the obtained wavefront branch is Expressed as:

[0099] ;

[0100] in, represents the convolution operation, is the dimension of the feature space, Represents the space of real vectors.

[0101] 3) MTF curve branch (Branch-M): Use one-dimensional convolutional neural network (1D-CNN) or frequency domain attention network to process MTF curve The frequency distribution characteristics of the MTF curve branch are obtained by Expressed as:

[0102] ;

[0103] in, Represents the dimension, represents the convolution operation, is the dimension of the feature space, Represents the space of real vectors.

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

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

[0106] ;

[0107] in, Indicates the The learnable weight vectors of the branches, Indicates the The feature vector of each branch is extracted by the branch network. Indicates the branch index of the current calculation, the value is , 、 、 Represent the learnable weight vectors of the Zernike branch, wavefront branch, and MTF curve branch, respectively. , , Represent the characteristic vectors of Zernike branch, wavefront branch, and MTF curve branch respectively, represents the vector transpose operation, Represents the normalized attention weight coefficient; this formula weights the features of each branch through the soft max operation to reflect the importance of each input in the final design. After weighting, the fusion feature vector It is given by:

[0108] ;

[0109] in, , , The eigenvectors are 、 and The normalized attention weight coefficient of ;

[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 and receives the fused high-dimensional fusion feature vector , which is mapped into the final optical design parameters through a multi-layer perceptron (MLP) Specifically, the prediction function is:

[0112] ;

[0113] in, are the predicted optical design parameters, is the mapping function of the multilayer perceptron network. The output optical design parameters include, but are not limited to, the curvature radius of each lens, lens thickness, lens refractive index, aperture position, total length of the optical system, field of view, and other key structural parameters. Through the predictions of this module, the network model can directly generate design parameters that meet the given optical performance objectives, avoiding the complex reverse optimization process required by traditional methods.

[0114] 2. Model training phase.

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

[0116] 1) Collect optical system design samples and construct a data set (i.e., training set), where: are known design parameters;

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

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

[0119] ;

[0120] in, are network model parameters, is the regularization 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) The network model is evaluated on the validation set for prediction accuracy and generalization ability until convergence.

[0123] 3. Model reasoning stage.

[0124] Model inference phase see Figure 2 (The figure only shows the outline of the steps), including the following steps:

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

[0126] 2) Load the network model parameters, and the performance indicators processed by the input feature construction module enter the network model and undergo a forward propagation;

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

[0128] 4) The optical design parameters can be It can be used directly for preliminary optical system construction or as an initial solution for fine-tuning in traditional optical design software.

[0129] Figure 2 The training and inference processes for the network model are demonstrated. Model training is first completed by collecting samples, constructing a dataset, building a network, defining a loss function, and optimizing parameters. The target performance metric is then input, the model is loaded, and forward propagation is performed to obtain the design parameters, thereby enabling the automated 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 invention can not only improve the efficiency and intelligence 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 broad application prospects and engineering value.

[0131] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A forward design method for an optical system based on multi-branch weighted feature fusion, characterized in that: The following steps are involved: Step 1: Selection and construction of input features; Taking the performance indicators of multiple optical systems as input features; Step 2: Multi-branch structure design; A multi-branch deep learning network structure is used, in which each branch is responsible for extracting optical design features from different input feature dimensions or scales; Step 3: Adaptive weighted feature fusion mechanism; Adopting an adaptive weighted feature fusion mechanism, the weight of each input feature is automatically adjusted according to its contribution to the final design parameters, thus achieving accurate prediction of optical system design parameters. Step 4: Forecast output of forward design; Using forward design, the corresponding optical system design parameters are directly predicted from multiple optical performance indicators; Step 5: Training and optimization process; During 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 through the back-propagation algorithm to achieve efficient performance of the multi-branch deep learning network structure model in the designed task; Step 6: Verification and application of output results; The trained multi-branch deep learning network structure model predicts the optical system design parameters in real time by inputting different performance indicators; The output design parameters are used in the actual manufacturing and debugging of the optical system, or provide a reference for optimization strategy during the design iteration process.

2. The optical system forward design method based on multi-branch weighted feature fusion according to claim 1, characterized in that: The performance indicators of the multiple optical systems in step 1 include: Zernike aberration coefficients, wavefront distribution diagrams, 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 of different optical systems and corresponding performance indicators.

4. A network model for the optical system forward design method based on multi-branch weighted feature fusion according to any one of claims 1 to 3, characterized in that: include: Input feature construction module, multi-branch feature extraction module, adaptive weighted feature fusion module and optical design parameter prediction module; wherein: The input feature construction module is used to receive multiple performance indicators of the optical system and preprocess them into input features of the network model; The multi-branch feature extraction module includes: Zernike branch, wavefront branch and MTF curve branch; among them: Zernike branch uses multi-layer perceptron to extract Zernike aberration coefficient Processing is performed to extract the global design features; the wavefront image branch uses a multi-layer convolutional neural network to process the wavefront distribution image , extract its spatial distortion pattern; the MTF curve branch uses a one-dimensional convolutional neural network or a frequency domain attention network to process the MTF curve Frequency distribution characteristics of The adaptive weighted feature fusion module is used to dynamically fuse the features extracted from the Zernike branch, wavefront branch, and MTF curve branch, taking into account the impact of different input sources on the final design; The optical design parameter prediction module is used to receive the fused high-dimensional features and map them into the final optical design parameters through a multi-layer perceptron.

5. The network model according to claim 4, characterized in that In the Zernike branch of the multi-branch feature extraction module, the feature vector of the Zernike branch It is given by the following formula: ； in, is the dimension of the feature space, represents a multilayer perceptron, represents the space of real vectors; In the wavefront branch of the multi-branch feature extraction module, the feature vector of the wavefront branch is obtained Expressed as: ； in, represents the convolution operation, is the dimension of the feature space, represents the space of real vectors; In the MTF curve branch of the multi-branch feature extraction module, the feature vector of the MTF curve branch is obtained Expressed as: ； in, Represents the dimension, represents the convolution operation, is the dimension of the feature space, Represents the space of real vectors.

6. The network model according to claim 4, characterized in that The adaptive weighted feature fusion module dynamically fuses features, specifically: Calculate the weight coefficient of each eigenvector ,in, , the weight coefficient of each branch feature vector is calculated by the following formula: ； in, Indicates the The learnable weight vectors of the branches, Indicates the The feature vector of each branch, Indicates the branch index of the current calculation, the value is , 、 、 Represent the learnable weight vectors of the Zernike branch, wavefront branch, and MTF curve branch, respectively. , , Represent the characteristic vectors of Zernike branch, wavefront branch, and MTF curve branch respectively, represents the vector transpose operation, Represents the normalized attention weight coefficient; After weighting, the fusion feature vector It is given by: ； in, , , The eigenvectors are 、 and The normalized attention weight coefficient.

7. The network model according to claim 6, characterized in that The prediction function of the optical design parameter prediction module is: ； in, are the predicted optical design parameters, is the mapping function of the multilayer perceptron network.

8. The network model according to claim 4, characterized in that The training process in its usage process includes the following steps: Step 1: Collect optical system design samples and construct a data set, where: are known design parameters; Step 2: Build the network architecture of the network model, take the performance indicators as input, and the network model outputs the prediction parameters ; Step 3: Define the loss function To optimize: ； in, are network model parameters, is the regularization coefficient, represents the sample index, Represents the total number of training samples; Step 4: Use the optimizer to perform gradient descent and update the network model parameters; Step 5: The network model evaluates prediction accuracy and generalization ability on the validation set until convergence.

9. The network model according to claim 4, characterized in that The model inference phase of its usage process includes the following steps: Step 1: Input the target optical system performance index combination; Step 2: After being processed by the input feature construction module, it enters the network model and undergoes one forward propagation; Step 3: The network model quickly outputs a set of optical design parameters ; Step 4: Optical design parameters It can be used directly for preliminary optical system construction or as an initial solution for fine-tuning in optical design software.

Images