An optical system initial structure determination method, a computer storage medium and an apparatus

Abstract

This invention relates to optical system design methods, computer storage media, and devices, specifically to a method, computer storage media, and device for determining the initial structure of an optical system. The aim is to address the problems of cumbersome and lengthy calculation processes, large accumulated errors, or low design efficiency in existing methods for obtaining the initial structure of an optical system. The method includes: calculating the design parameters of the optical system based on design requirements; using a particle swarm optimization algorithm to obtain multiple initial optical system structures as reference designs; normalizing the structural parameters of the reference designs; training a deep neural network model using a weighted hybrid loss function of supervised and unsupervised loss functions; inputting the design parameters of the optical system to be designed into the trained network model, and outputting the initial structure parameters. This invention combines particle swarm optimization and deep learning, enabling rapid and accurate generation of the initial structure of an optical system, significantly improving design efficiency.

Description

Technical Field

[0001] This invention relates to optical system design methods, computer storage media and devices, and more specifically to a method for determining the initial structure of an optical system, a computer storage medium and devices. Background Technology

[0002] Currently, optical engineers primarily use two methods to obtain the initial structure of an optical system during the design process. One method is the analytical method, which calculates the structural parameters that meet the system requirements based on the primary aberration theory of the optical system, serving as the initial structure. However, this method is only suitable for simple optical systems; for complex systems, the calculation process is cumbersome and lengthy, accumulating significant errors, and therefore it is rarely used in engineering practice. The other method is the scaling method, which involves searching for an optical system similar to the design requirements in existing patent literature or lens databases as the initial structure and scaling its focal length. This method is currently used by most optical engineers, but the design result of the optical system is highly dependent on the acquisition of the existing initial structure. If the initial structure is not properly obtained, subsequent optimization will be difficult to achieve ideal performance, resulting in low design efficiency.

[0003] Chinese patent CN116009246A discloses an automatic optimization design method for polarization optical systems based on deep learning, which includes the following steps: 1) generating a normalized optical system sample dataset; 2) training a deep neural network model; standardizing the output of the established deep neural network, the output of which is the curvature, thickness variable, refractive index of the optical glass, and Abbe number of the optical system, wherein the curvature, glass refractive index, and Abbe number remain unchanged, and only the original thickness variable is standardized; 3) calculating unsupervised loss by ray tracing and supervising loss by combining the standardized results with the labeled data; 4) performing weighted calculation by combining the supervised loss and unsupervised loss; 5) updating network parameters through backpropagation; determining whether the iteration is complete, if the iteration is not complete, repeating the previous steps until the iteration is complete, if the iteration is complete, inputting the indicators into the deep neural network to automatically optimize the design of the polarization optical system. In this design method, the reference lenses are only from a fixed lens library, the sample coverage is limited, and the process economy of lens materials and the failure and surface overlap constraints in ray tracing are not considered.

[0004] Chinese patent CN115963628A discloses a deep learning optical imaging system design method based on complex function neural networks, which includes the following steps: 1) Taking the object distance, system focal length, and F-number as inputs, and determining the image distance and imaging scaling ratio according to the imaging relationship of an ideal optical system; 2) Calculating the image-side image of the ideal optical system based on the imaging scaling ratio and the object-side image of the ideal optical system; 3) Constructing a complex function generator network model; 4) Constructing a discriminator network model; 5) Completing the establishment and training of the above-mentioned deep learning model of the complex function neural network; 6) The usage stage. This optical imaging design method does not require an initial structure, but relies on generative adversarial networks for image similarity discrimination, and does not solve the problem of how to directly generate a physically reasonable initial structure that can be directly used for subsequent optimization.

[0005] Therefore, in order to solve the problem of the difficulty in accurately obtaining the initial structure in current optical design, there is an urgent need for a method that can quickly and accurately obtain the initial structure of the optical system in order to improve the design efficiency of the optical system. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of cumbersome and lengthy calculation processes, large cumulative errors, or low design efficiency in existing methods for obtaining the initial structure of an optical system, and to provide a method, computer storage medium, and device for determining the initial structure of an optical system.

[0007] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0008] A method for determining the initial structure of an optical system, characterized by the following steps:

[0009] Step 1: Calculate the design specifications of the optical system based on its design requirements;

[0010] Step 2: Using the design specifications of the optical system as input, the particle swarm optimization algorithm is used to obtain multiple initial structures of the optical system as reference designs.

[0011] Step 3: Construct a deep neural network model, normalize the structural parameters of the reference design, expand the number of samples according to the reference design, and then train the deep neural network model using a weighted mixture of supervised and unsupervised loss functions to obtain the trained deep neural network model.

[0012] Step 4: Input the design specifications of the optical system to be designed into the trained deep neural network model. The deep neural network model outputs the initial structural parameters of the optical system, thus completing the determination of the initial structure of the optical system.

[0013] Furthermore, in step 1, the design specifications of the optical system include the working band, focal length, field of view, aperture, number of lenses N, object distance, image distance, total length, and imaging quality requirements of the optical system, where N is a positive integer.

[0014] Furthermore, step 2 specifically involves:

[0015] Step 2.1: Based on the number of lenses N and focal length requirements of the optical system, use the optical power allocation evaluation function, which includes the optical power formula and the achromatic formula. Perform the initial allocation of optical power for the optical system to obtain the optical power allocation results for each lens;

[0016] Step 2.2: Based on the hardness, wear resistance, and price of each material in the existing optical glass material library, use the material matching evaluation function. The optimal optical material for each lens is matched to the one with the best process economy, and the lens material matching results are obtained.

[0017] Step 2.3: Combining the power allocation results of each lens and the lens material matching results, use the comprehensive evaluation function. The radii of curvature of the front and rear surfaces of each lens are obtained as the initial structure of an optical system.

[0018] Step 2.4: Change the focal length of the optical system and obtain a total of Z initial structures of the optical system as reference designs, following the methods in Steps 2.1 to 2.3.

[0019] Furthermore, step 3 specifically involves:

[0020] Step 3.1: Construct and initialize the deep neural network model;

[0021] Step 3.2: Determine the input and output of the deep neural network model, where the input is the design specifications of the optical system and the output is the initial structural parameters of the optical system;

[0022] Step 3.3: Normalize the structural parameters of the reference design, increase the number of samples, and generate supervised training samples and unsupervised training samples;

[0023] Step 3.4: Calculate the mean square error between the structural parameters of the reference design and the structural parameters output by the deep neural network model, and establish the supervised loss function. ;

[0024] Step 3.5: Establish a differentiable ray tracing module, calculate the spot size of the optical system output by the deep neural network model, and establish an unsupervised loss function. ;

[0025] Step 3.6: The supervised loss function... With unsupervised loss function We perform weighted mixing to obtain the mixed loss function. Then, based on the mixed loss function Batch training is performed on the deep neural network model to obtain the trained deep neural network model.

[0026] Further, in step 2.1, N=2, including a first lens and a second lens;

[0027] The optical power allocation evaluation function for:

[0028]

[0029] in, This is a weighting coefficient for the optical power of the optical system. This is the weighting coefficient for chromatic aberration in the optical system. The total optical power is calculated using the particle swarm optimization algorithm. This represents the total optical power required by the optical system, with the subscript 'i' indicating the lens number. Let be the Abbe number of the i-th lens. Let i be the optical power of the i-th lens;

[0030] In step 2.2, the material matching evaluation function for:

[0031]

[0032] in, This is a weighting coefficient for the economic efficiency of materials and processes in optical systems. Let be the process economic index for the i-th lens;

[0033] The process economy is determined by the process economy index C, and its calculation formula is as follows:

[0034] ;

[0035] in, The hardness of the lens material, The wear degree of the lens material. For the price of the lens material, ω HK ω is the hardness weighting coefficient for the lens material. FA ω is the weighting coefficient for lens material wear. PR The weighting coefficient for lens material price;

[0036] In step 2.3, the comprehensive evaluation function for:

[0037] ;

[0038] Where, n i Let r be the refractive index of the i-th lens material. 2i-1 Let r be the radius of curvature of the front surface of the i-th lens. 2i Let be the radius of curvature of the back surface of the i-th lens.

[0039] Furthermore, in step 3.1, the deep neural network model includes an input layer, seven hidden layers, and an output layer;

[0040] In step 3.3, the number of supervised training samples and the number of unsupervised training samples are both 1024;

[0041] In step 3.4, the supervised loss function for:

[0042] ;

[0043] Where M is the total number of lens surfaces, M=4; j is the lens surface number, where For the curvature of the j-th lens surface in the reference design, Let be the curvature of the j-th lens surface output by the deep neural network model. For reference, the thickness of the i-th lens in the design, Let be the thickness of the i-th lens output by the deep neural network model. For reference, the refractive index of the i-th lens in the design, Let be the refractive index of the i-th lens output by the deep neural network model. For reference, the Abbe number of the i-th lens in the design, The Abbe number of the i-th lens output by the deep neural network model;

[0044] In step 3.5, the unsupervised loss function ;in, N represents the weight of the penalty function. H N represents the number of fields of view. W N represents the number of wavelengths. p q represents the number of light rays entering the entrance pupil from the same field of view and with the same wavelength. Hwp Here, is the penalty function, and s is the RMS value of the optical system output by the deep neural network model;

[0045] In step 3.6, the hybrid loss function ,in To monitor the weights of the loss function, Let be the average unsupervised loss function for batch training. This is the average supervised loss function for batch training.

[0046] Meanwhile, the present invention also provides a computer storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the above-described method for determining the initial structure of an optical system.

[0047] In addition, the present invention also provides a computer device, including a processor, a memory connected to the processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for determining the initial structure of an optical system.

[0048] Compared with the prior art, the present invention has the following beneficial technical effects:

[0049] 1. The initial structure determination method for optical systems provided by this invention improves the design efficiency of refractive optical systems and overcomes the limitations of existing optical design software algorithms. Specifically, this invention employs a technical approach combining particle swarm optimization and deep neural networks, enabling the rapid generation of initial structures for optical systems that meet design specifications without the need for repeated manual trial and error. Compared to the cumbersome process of traditional scaling or analytical calculations, this invention significantly shortens the initial structure acquisition time, lowers the design threshold, and effectively overcomes the algorithmic limitations of existing commercial optical design software (such as Zemax and CodeV) in the automatic generation of initial structures.

[0050] 2. The initial structure determination method for optical systems provided by this invention addresses the problem of limited training samples in deep learning for optical systems by employing a hybrid supervised and unsupervised learning optimization approach, while further improving the image quality of the initial structure. The supervised learning method uses multiple reference designs generated by the particle swarm optimization algorithm as labels to ensure the physical rationality of the network's output parameters. The unsupervised learning method introduces a differentiable ray tracing module, using the spot size (RMS) as an evaluation metric, enabling the network to autonomously optimize imaging quality. This hybrid training strategy effectively solves the problem of limited training samples in deep learning for optical systems, while also ensuring that the generated initial structure has superior imaging quality compared to traditional methods, providing a better starting point for subsequent fine-tuning.

[0051] 3. The initial structure determination method for optical systems provided by this invention can be applied not only to the design of refractive optical systems, but also extended to the design of reflective optical systems by changing the differentiable ray tracing module and evaluation function. Therefore, the method of this invention is not only applicable to refractive optical systems (such as visible light lenses, lithography objectives, etc.), but can also be directly extended to the initial structure design of reflective optical systems (such as telescopes, satellite optical payloads, etc.) by replacing the differentiable ray tracing module and adjusting the evaluation function. This versatility gives this invention cross-domain application value, providing a unified technical framework for the automated design of different types of optical systems. Attached Figure Description

[0052] Figure 1 This is a flowchart of an embodiment of the method for determining the initial structure of the optical system according to the present invention;

[0053] Figure 2 This is a graph showing the loss function of the deep neural network model training in an embodiment of the method for determining the initial structure of the optical system of the present invention.

[0054] Figure 3 This is a schematic diagram and dot plot of a reference design with the same focal length as the system to be designed in an embodiment of the method for determining the initial structure of the optical system of the present invention;

[0055] Figure 4 This is a schematic diagram and dot plot of the predictive design in an embodiment of the method for determining the initial structure of the optical system of the present invention. Detailed Implementation

[0056] To make the objectives, advantages, and features of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0057] A method for determining the initial structure of an optical system, the flowchart of which is shown below. Figure 1 This includes the following steps:

[0058] Step 1: Calculate the design specifications of the optical system based on the design requirements of the optical system. The design specifications of the optical system include the working band, focal length, field of view, aperture, number of lenses N, object distance, image distance, total length, and imaging quality requirements.

[0059] Step 2: Using the design specifications of the optical system as input, the particle swarm optimization algorithm is used to obtain multiple initial structures of the optical system as reference designs.

[0060] Step 2.1: Based on the design requirements of the optical system and combined with engineering experience, determine the number of lenses N required for the optical system. In this embodiment, N=2, including the first lens and the second lens; then, based on the focal length requirements of the optical system, use the optical power allocation evaluation function that includes the optical power formula and the achromatic formula. Perform the initial allocation of optical power for the optical system:

[0061]

[0062] in, This is a weighting coefficient for the optical power of the optical system. This is the weighting coefficient for chromatic aberration in the optical system. The total optical power is calculated using the particle swarm optimization algorithm. This represents the total optical power required by the optical system, with the subscript 'i' indicating the lens number. Let be the Abbe number of the i-th lens. Let be the optical power of the i-th lens. The optical power of the first lens. The optical power of the second lens. Let be the Abbe number of the first lens. Let be the Abbe number of the second lens; where is the weighting coefficient for the chromatic aberration of the optical system. The priority of chromatic aberration correction of the optical system is taken as the core, and is determined by combining the system's spectral band, aberration index and application scenario.

[0063] Step 2.2: Based on the hardness, wear resistance, and price of each material in the existing optical glass material library, use the material matching evaluation function F... mat Matching the most economical optical material to each lens:

[0064]

[0065] in, This is a weighting coefficient for the economic efficiency of materials and processes in optical systems. Let be the economic performance index of the i-th lens. The economic indicators for the manufacturing process of the first lens are as follows: This is a manufacturing economy indicator for the second lens. The weighting coefficient for the material manufacturing economy of the optical system is... The core considerations are cost control objectives, processing difficulty, and mass production requirements, along with the weighting coefficients for chromatic aberration in the optical system. To achieve a balance between optical performance and process economy.

[0066] The economic efficiency of a process is determined by the economic efficiency index C, and its calculation formula is as follows:

[0067] ;

[0068] in, The hardness of the lens material, The wear degree of the lens material. For the price of the lens material, ω HK ω is the hardness weighting coefficient for the lens material. FA ω is the weighting coefficient for lens material wear. PR This represents the weighting coefficient for the lens material price. Among them, the lens material hardness weighting coefficient ω... HK Based on the lens's manufacturing process requirements, operating environment, and durability needs; the lens material wear coefficient ω FA The weighting coefficient ω is based on the difficulty of lens processing, yield rate, and surface quality requirements; the lens material price is also considered. PR Based on project budget, mass production scale and material procurement costs, the cost ratio of optical materials is controlled, while taking into account the stability of material supply and the risk of price increases for rare materials.

[0069] Table 1 lists the performance parameters of some glass materials produced by Chengdu Guangming Optoelectronic Co., Ltd.

[0070] Table 1 Performance parameters of some glass materials produced by Chengdu Guangming Optoelectronic Co., Ltd.

[0071]

[0072] Step 2.3: Combining the power allocation results of each lens and the lens material matching results, use the comprehensive evaluation function. The radii of curvature of the front and rear surfaces of each lens are obtained as the initial structure of an optical system.

[0073] ;

[0074] Where, r 2i-1 Let r be the radius of curvature of the front surface of the i-th lens. 2i Let r1 be the radius of curvature of the rear surface of the i-th lens, n1 be the refractive index of the first lens material, n2 be the refractive index of the second lens material, r1 be the radius of curvature of the front surface of the first lens, r2 be the radius of curvature of the rear surface of the first lens, r3 be the radius of curvature of the front surface of the second lens, and r4 be the radius of curvature of the rear surface of the second lens. This refers to the total optical power required by the optical system.

[0075] Step 2.4: Change the focal length of the optical system and obtain the initial structures of ten optical systems in total, following the methods in Steps 2.1 to 2.3, as reference designs.

[0076] Step 3: Normalize the structural parameters of the reference design, expand the number of samples according to the reference design, and then train the deep neural network model using a weighted mixture of supervised and unsupervised loss functions to obtain the trained deep neural network model.

[0077] Step 3.1: Construct and initialize the deep neural network model. Based on the highly nonlinear and strongly coupled mapping relationship between optical system design indicators and initial structural parameters, this embodiment uses a stacked fully connected deep neural network model as the modeling tool. This deep neural network model consists of an input layer, several hidden layers, and an output layer. A nonlinear activation function is introduced after each fully connected layer to enhance the model's ability to express complex nonlinear relationships. In the structural design of the deep neural network model, a single sub-network contains seven hidden layers, each with thirty-two neurons. To improve the model's predictive stability and generalization ability, sixteen sub-networks with identical structures and independently initialized parameters are constructed. The output results of each sub-network are averaged as the final prediction result.

[0078] Step 3.2: Determine the input and output of the deep neural network model, where the input is the design specifications of the optical system and the output is the initial structural parameters of the optical system;

[0079] Step 3.3: Normalize the structural parameters of the reference design to make training faster and more stable, expand the number of samples, and generate 1024 supervised training samples and 1024 unsupervised training samples.

[0080] Step 3.4: Calculate the mean square error between the structural parameters of the reference design and the structural parameters output by the deep neural network model, and establish the supervised loss function. ; ;

[0081] Where M is the total number of lens surfaces, and in this embodiment M=4; j is the lens surface number, where For the curvature of the j-th lens surface in the reference design, Let be the curvature of the j-th lens surface output by the deep neural network model. For reference, the thickness of the i-th lens in the design, Let be the thickness of the i-th lens output by the deep neural network model. For reference, the refractive index of the i-th lens in the design, Let be the refractive index of the i-th lens output by the deep neural network model. For reference, the Abbe number of the i-th lens in the design, This is the Abbe number of the i-th lens output by the deep neural network model.

[0082] Step 3.5: Establish a differentiable ray tracing module, calculate the spot size of the optical system output by the deep neural network model, and establish an unsupervised loss function. Unsupervised learning training is unlabeled training that requires a custom loss function. Supervised training only learns the reference design numerically but does not constrain the system's imaging quality. In evaluating the imaging quality of optical systems, the system's spot size RMS (Root Mean Square) is usually used as an evaluation metric. To calculate the RMS, differentiable ray tracing is performed using the structural parameters and design specifications output by DNNs (Deep Neural Networks), including the radius of curvature, optical surface thickness, refractive index of the glass material, and Abbe number. The formula for calculating RMS is:

[0083]

[0084] Where, N H N represents the number of fields of view, typically 0, 0.7, and 1. H =3;N W This indicates the number of wavelengths. In this embodiment, the wavelengths are visible light F, d, C; N P This represents the number of rays entering the entrance pupil under the same field of view and wavelength. A larger value results in a more accurate calculated RMS value, but also increases computational complexity and reduces model training speed. In this embodiment, 17 is used. HWP The image height at a given field of view, wavelength, and entrance pupil diameter. is the average image height; s is the RMS value of the optical system output by the deep neural network model.

[0085] Meanwhile, to constrain ray failures and optical surface overlaps during ray tracing, a penalty function is defined. :

[0086] ;

[0087] Where, N k Let I be the number of light rays, k be the interface encountered by the light rays, and I be the number of light rays. Hwpk Let I' be the angle of incidence of the light ray on interface k. Hwpk Let Δz be the angle of incidence of the ray at interface k. Hwpk This represents the displacement of light rays between adjacent surfaces.

[0088] The unsupervised loss function is a combination of the spot size and the penalty function, as shown in the following formula:

[0089]

[0090] Where, λ qThe weight of the penalty function is 10 in this embodiment. 3 ;

[0091] Step 3.6: The supervised loss function... With unsupervised loss function We perform weighted mixing to obtain the mixed loss function. Then, based on the mixed loss function Batch training is performed on the deep neural network model to obtain the trained deep neural network model. Where λ s To monitor the weights of the loss function, a value of 10 is used in this embodiment; Let be the average unsupervised loss function for batch training. This is the average supervised loss function for batch training.

[0092] Step 4: Input the design specifications of the optical system to be designed into the deep neural network model. The deep neural network model outputs the initial structural parameters of the optical system, completing the determination of the initial structure of the optical system. Taking the two-piece lens in this embodiment as an example, the design specifications of the optical system are:

[0093] Operating wavelength: 486~656nm, focal length: 100mm, aperture: 20mm, field of view: 4°.

[0094] Table 2 shows the structural parameters of a reference design with the same focal length as the system to be designed.

[0095] Table 2. Structural parameters of a reference design with the same focal length as the system under design.

[0096]

[0097] Table 3 shows the initial structural parameters of the optical system.

[0098] Table 3 Initial structural parameters of the optical system

[0099]

[0100] Depend on Figure 2 As can be seen, the supervised loss function curve (blue), the unsupervised loss function curve (red), and the total loss function curve (green) all show a steady downward trend and eventually converge, proving the effectiveness and stability of the hybrid training strategy. Combined with... Figure 3 and Figure 4 It can be seen that the RMS value of the point plot of the output design is comparable to that of the point plot of the reference design, which verifies the effectiveness of the model.

[0101] In addition, this embodiment also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for determining the initial structure of an optical system.

[0102] In addition, this embodiment also provides a computer device, including a processor, a memory connected to the processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for determining the initial structure of an optical system.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A method for determining the initial structure of an optical system, characterized in that, Includes the following steps: Step 1: Calculate the design specifications of the optical system based on its design requirements; Step 2: Using the design specifications of the optical system as input, the particle swarm optimization algorithm is used to obtain multiple initial structures of the optical system as reference designs. Step 3: Construct a deep neural network model, normalize the structural parameters of the reference design, expand the number of samples according to the reference design, and then train the deep neural network model using a weighted mixture of supervised and unsupervised loss functions to obtain the trained deep neural network model. Step 4: Input the design specifications of the optical system to be designed into the trained deep neural network model. The deep neural network model outputs the initial structural parameters of the optical system, thus completing the determination of the initial structure of the optical system.

2. The method for determining the initial structure of an optical system according to claim 1, characterized in that: In step 1, the design specifications of the optical system include the working band, focal length, field of view, aperture, number of lenses N, object distance, image distance, total length, and imaging quality requirements, where N is a positive integer.

3. The method for determining the initial structure of an optical system according to claim 2, characterized in that: Step 2 is as follows: Step 2.1: Based on the number of lenses N and focal length requirements of the optical system, use the optical power allocation evaluation function, which includes the optical power formula and the achromatic formula. Perform the initial allocation of optical power for the optical system to obtain the optical power allocation results for each lens; Step 2.2: Based on the hardness, wear resistance, and price of each material in the existing optical glass material library, use the material matching evaluation function. The optimal optical material for each lens is matched to the one with the best process economy, and the lens material matching results are obtained. Step 2.3: Combining the power allocation results of each lens and the lens material matching results, use the comprehensive evaluation function. The radii of curvature of the front and rear surfaces of each lens are obtained as the initial structure of an optical system. Step 2.4: Change the focal length of the optical system and obtain a total of Z initial structures of the optical system as reference designs, following the methods in Steps 2.1 to 2.

3.

4. The method for determining the initial structure of an optical system according to claim 3, characterized in that: Step 3 specifically involves: Step 3.1: Construct and initialize the deep neural network model; Step 3.2: Determine the input and output of the deep neural network model, where the input is the design specifications of the optical system and the output is the initial structural parameters of the optical system; Step 3.3: Normalize the structural parameters of the reference design, increase the number of samples, and generate supervised training samples and unsupervised training samples; Step 3.4: Calculate the mean square error between the structural parameters of the reference design and the structural parameters output by the deep neural network model, and establish the supervised loss function. ; Step 3.5: Establish a differentiable ray tracing module, calculate the spot size of the optical system output by the deep neural network model, and establish an unsupervised loss function. ; Step 3.6: The supervised loss function... With unsupervised loss function We perform weighted mixing to obtain the mixed loss function. Then, based on the mixed loss function Batch training is performed on the deep neural network model to obtain the trained deep neural network model.

5. The method for determining the initial structure of an optical system according to claim 4, characterized in that: In step 2.1, N=2, including a first lens and a second lens; The optical power allocation evaluation function for: ； in, This is a weighting coefficient for the optical power of the optical system. This is the weighting coefficient for chromatic aberration in the optical system. The total optical power is calculated using the particle swarm optimization algorithm. This represents the total optical power required by the optical system, with the subscript 'i' indicating the lens number. Let be the Abbe number of the i-th lens. Let i be the optical power of the i-th lens; In step 2.2, the material matching evaluation function for: ； in, This is a weighting coefficient for the economic efficiency of materials and processes in optical systems. Let be the process economic index for the i-th lens; The process economy is determined by the process economy index C, and its calculation formula is as follows: ； in, The hardness of the lens material, The wear degree of the lens material. For the price of the lens material, ω HK ω is the hardness weighting coefficient for the lens material. FA ω is the weighting coefficient for lens material wear. PR The weighting coefficient for lens material price; In step 2.3, the comprehensive evaluation function for: ； Where, n i Let r be the refractive index of the i-th lens material. 2i-1 Let r be the radius of curvature of the front surface of the i-th lens. 2i Let be the radius of curvature of the back surface of the i-th lens.

6. The method for determining the initial structure of an optical system according to claim 5, characterized in that: In step 3.1, the deep neural network model includes one input layer, seven hidden layers, and one output layer; In step 3.3, the number of supervised training samples and the number of unsupervised training samples are both 1024; In step 3.4, the supervised loss function for: ； Where M is the total number of lens surfaces, M=4; j is the lens surface number, where For the curvature of the j-th lens surface in the reference design, Let be the curvature of the j-th lens surface output by the deep neural network model. For reference, the thickness of the i-th lens in the design, Let be the thickness of the i-th lens output by the deep neural network model. For reference, the refractive index of the i-th lens in the design, Let be the refractive index of the i-th lens output by the deep neural network model. For reference, the Abbe number of the i-th lens in the design, The Abbe number of the i-th lens output by the deep neural network model; In step 3.5, the unsupervised loss function ;in, N represents the weight of the penalty function. H N represents the number of fields of view. W N represents the number of wavelengths. p q represents the number of light rays entering the entrance pupil from the same field of view and with the same wavelength. Hwp Here, is the penalty function, and s is the RMS value of the optical system output by the deep neural network model; In step 3.6, the hybrid loss function ,in To monitor the weights of the loss function, Let be the average unsupervised loss function for batch training. This is the average supervised loss function for batch training.

7. A computer storage medium storing a computer program thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for determining the initial structure of the optical system according to any one of claims 1-6.

8. A computer device comprising a processor, a memory connected to the processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the method for determining the initial structure of the optical system according to any one of claims 1-6.

Images