Lens design method based on deep learning

Abstract

This invention discloses a lens design method based on deep learning, belonging to the fields of optical design and deep learning. The method includes: constructing a dual-channel neural network; adjusting task indicators in real time according to the current state using a dynamic task generator; inputting the task indicators into the dual-channel neural network for joint training; calculating aberration sensitivity based on Zernike aberration coefficients from real-time ray tracing verification feedback; and outputting a lens design result that meets both optical performance and manufacturing feasibility requirements when the closed-loop iteration formed by dual-channel joint training, real-time ray tracing verification, and dynamic optimization strategy adjustment satisfies a preset convergence condition. This invention uses a dynamic task generator to adjust the optimization path in real time, physically constrains channels to block unmanufacturable structures, and prioritizes suppressing dominant aberrations through aberration sensitivity feedback; it embeds lightweight ray tracing into the training loop, provides real-time feedback of Zernike aberration coefficients, and uses manufacturing feasibility scoring to drive design iteration.

Description

Technical Field

[0001] This invention relates to optical design and deep learning technology, and particularly to a lens design method based on deep learning. Background Technology

[0002] Lens design is a crucial aspect of optics, aiming to determine the structural parameters of a lens, including the radius of curvature, thickness, material, and spacing between individual lens elements, based on specific application requirements such as image quality, focal length, and aperture size, in order to achieve the desired optical performance.

[0003] Existing technologies have disclosed lens design methods utilizing deep learning. For example, Chinese patent application CN107976804 A discloses a design method, device, equipment, and storage medium for a lens optical system; CN116187162A discloses a model-driven deep learning-based large field-of-view multi-metasurface lens design method. However, traditional lens design methods have the following drawbacks:

[0004] (1) Low efficiency: Traditional lens design methods often rely on a lot of repetitive manual calculations and simulations. Designers need to constantly adjust parameters and evaluate results. The whole process is time-consuming and laborious. Especially for complex optical systems, the design cycle may last for months or even years, which is difficult to meet the needs of rapid iteration in modern industry.

[0005] (2) Traditional methods rely heavily on the experience of designers: designers need to have in-depth optical knowledge and rich practical experience in order to make reasonable parameter adjustments. However, experience is subjective and limited. Differences in the thinking and level of different designers may lead to inconsistent design results, making it difficult to guarantee the consistency and stability of the design.

[0006] (3) In addition, traditional design methods sometimes deviate from physical constraints during the optimization process: in order to pursue certain performance indicators, unreasonable structural parameters may be generated, making it difficult to realize the design in actual manufacturing, increasing the risk and cost of design failure. Summary of the Invention

[0007] Purpose of the invention: To address the above problems, the purpose of this invention is to provide a lens design method based on deep learning.

[0008] Technical solution: The deep learning-based lens design method of the present invention includes:

[0009] Construct a two-channel neural network for lens design, wherein the two-channel neural network includes a performance prediction channel and a physical constraint channel;

[0010] The dynamic task generator adjusts task metrics in real time based on the current status.

[0011] The task metrics output by the dynamic task generator are input into the dual-channel neural network for joint training, and the optical parameters are obtained based on the trained dual-channel neural network.

[0012] Lightweight ray tracing is embedded during training, and Zernike aberration coefficients are extracted based on the tracing results as real-time feedback signals to verify the optical performance of the dynamic task generator output.

[0013] Aberration sensitivity is calculated based on Zernike aberration coefficients from real-time ray tracing verification feedback. Dominant aberrations are suppressed first based on aberration sensitivity, and the weights of the optimization strategy are dynamically adjusted based on manufacturing feasibility feedback from the physical constraint channel output.

[0014] When the closed-loop iteration formed by dual-channel joint training, real-time ray tracing verification, and dynamic optimization strategy adjustment meets the preset convergence conditions, the output lens design result satisfies both optical performance and manufacturing feasibility requirements.

[0015] Preferably, the steps of adjusting task metrics in real time based on the current state using a dynamic task generator include:

[0016] Step 21, record the current state as ,in, This represents the average root mean square spot radius. This represents the radius of the maximum root mean square spot. These represent the Zernike aberration coefficients for terms 3, 5, and 8, respectively; the RMS error distribution is obtained by tracing the feature rays.

[0017] Step 22: Analyze the aberrations to obtain the Zernike aberration coefficients of the dominant aberrations;

[0018] Step 23, calculate aberration sensitivity;

[0019] Step 24, perform policy network decision-making, the calculation formula is:

[0020] ,

[0021] In the formula, Represents the action value function. Indicates an action, This represents the set of trainable parameters for a Q-network.

[0022] Step 25, calculate the reward function, the formula is:

[0023] ,

[0024] In the formula, Indicates the reward value. , Indicates the weighting coefficient. This represents the full field-of-view correction factor. Indicates the actual field of view of the current camera. Indicates the target's field of view;

[0025] Step 26, score the physical feasibility. Make a judgment, if , If the physical feasibility score threshold is used, then the step size of the task indicators is adjusted to make... , ;

[0026] The updated task marker is as follows ,in, This indicates the updated field of view. This indicates the updated aperture number. Indicates the effective focal length.

[0027] Preferably, the step of inputting the task metrics output by the dynamic task generator into a dual-channel neural network for joint training includes:

[0028] Step 31: Input the current task metrics into the performance prediction channel to obtain the optical system structure parameter vector, represented as:

[0029] ,

[0030] In the formula, This represents the performance prediction channel model; This represents a vector of structural parameters of the optical system. Indicates the first The radius of curvature of the spherical / aspherical surface. This indicates the center thickness of the lens or the air gap between lenses. Indicates the first The type of optical glass used in block lenses;

[0031] Step 32: Input the optical system structural parameter vector into the physical constraint channel to obtain the physical feasibility score, expressed as:

[0032] ,

[0033] In the formula, Represents a physical constraint channel model;

[0034] Step 33, calculate the total loss, the formula is:

[0035] ,

[0036] In the formula, Indicates loss of optical performance. Indicates physical constraint loss, This represents the aberration sensitivity-weighted loss. The physical constraint loss weighting coefficient, This is the weighting coefficient for aberration sensitivity loss;

[0037] Step 34: Update the parameters of the two-channel neural network using the AdamW optimizer. The updated formula is:

[0038] ,

[0039] In the formula, This indicates the step size for each parameter update. This represents the gradient operator of the loss function with respect to the parameter θ.

[0040] Preferably, the step of dynamically adjusting the weights of the optimization strategy based on the manufacturing feasibility feedback output from the physical constraint channel includes:

[0041] Step 51, analyze the current RMS distribution, if Then adjust the aberration weights using the following formula:

[0042] ,

[0043] In the formula, This represents the RMS spot radius at the edge of the field of view. This represents the RMS spot radius of the central field of view. The optimized weights for coma are represented;

[0044] Step 52, score based on physical feasibility. Adaptive learning rate adjustment ,like Then adjust to: , ;like Then adjust to: ;

[0045] Step 53: Adjust the weighting coefficients of the total loss function.

[0046] Preferably, when the closed-loop iteration formed by dual-channel joint training, real-time ray tracing verification, and dynamic optimization strategy adjustment meets the preset convergence condition, the steps for outputting lens design results that meet both optical performance and manufacturing feasibility requirements include:

[0047] When physical feasibility score And maximum field of view Then it is considered that the convergence condition is met. Indicate the operating wavelength and calculate the tolerance sensitivity matrix. , represented as:

[0048]

[0049] Based on the tolerance sensitivity matrix The cost is estimated using the following formula:

[0050] ,

[0051] In the formula, For the total cost of the lens, For the first Lens material cost Weights are used to penalize complexity. This is the design complexity function calculated based on the tolerance sensitivity matrix;

[0052] Generate a lens design report, including project design specifications, core optical system parameters, image quality simulation analysis, manufacturing cost estimates, and final design achievement conclusions and optimization suggestions.

[0053] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0054] 1. This invention proposes a dynamic-physical collaborative optimization mechanism to achieve real-time balance of multiple objectives:

[0055] Based on the current network status and historical optimization trajectory, diverse training tasks are adaptively generated to avoid the problem of getting trapped in local optima by a single optimization path, significantly improving the global optimization capability. During the training process, compliance checks are performed in real time on the geometric parameters, material properties, and manufacturing tolerances of the lens structure to filter out designs that do not meet the preset manufacturing process capabilities, ensuring the engineering feasibility of the output solution. By calculating the sensitivity of each optical surface to different aberration types, the aberration components that have the greatest impact on image quality are dynamically identified and prioritized for suppression, achieving efficient allocation of limited design resources.

[0056] 2. This invention constructs a closed-loop verification system to achieve integrated iteration of design, verification, and feedback:

[0057] The optical performance is rapidly calculated in each training iteration, overcoming the technical obstacles of large computational load and difficulty in integration into the deep learning training process of traditional ray tracing. Wavefront aberration is quantified into Zernike coefficients, which are directly fed back to the deep learning network as differentiable feedback signals, enabling the network to learn the mapping relationship from structural parameters to optical performance end-to-end. A quantitative manufacturing feasibility score is generated by combining indicators such as lens manufacturability, assembly tolerance sensitivity, material cost and thermal stability, and this score is used as the driving signal for iterative optimization.

[0058] 3. This invention introduces a tolerance sensitivity matrix to further improve design accuracy:

[0059] Construct a tolerance sensitivity matrix to systematically evaluate the performance degradation of each design parameter under manufacturing tolerance fluctuations, thereby quantifying the manufacturing robustness of the design and ensuring that the output solution not only meets nominal performance indicators but also has good tolerance capabilities. Attached Figure Description

[0060] Figure 1 This is a flowchart of a deep learning-based lens design method. Detailed Implementation

[0061] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0062] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0063] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0064] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0065] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0066] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0067] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0068] Combination Figure 1 As shown in this embodiment, the deep learning-based lens design method includes the following steps:

[0069] Step 1: Construct a dual-channel neural network for lens design, which includes a performance prediction channel and a physical constraint channel.

[0070] In one example, a two-channel neural network is constructed for lens design, including a performance prediction channel and a physical constraint channel. Specifically, the performance prediction channel uses a ResNet-18 network to process the design metric tensor. The input is fed into the ResNet-18 network to obtain the optical parameter vector. ,in For the field of view, This refers to the aperture number. For the effective focal length, Let be the radius of curvature of the spherical / aspherical surface of the nth surface. This refers to the center thickness of the lens or the air gap between lenses. This specifies the optical glass type for each lens. The physical constraint channel employs a 3-layer MLP structure to vectorize the optical parameters. After inputting into the 3-layer MLP structure, a physical feasibility score is obtained. The calculation formula is:

[0071] ,

[0072] In the formula, It is the Sigmoid activation function. , , The weight matrix for each layer of the network. , , These are the bias vectors for each network layer. It is a rectified linear activation function.

[0073] Step 2: Adjust task metrics in real time based on the current status using a dynamic task generator.

[0074] Specifically, a reinforcement learning policy network is adopted as the core structure of the dynamic task generator. Taking the current system state as input, it outputs dynamically optimized actions and updated design task metrics, achieving adaptive adjustment of the task objective. The reinforcement learning policy network is not used directly; firstly, the network weights and biases are randomly initialized without separate offline pre-training. During the system's closed-loop iteration process, the current imaging state, aberration distribution, and physical feasibility score are used as input. The network parameters are updated online based on the optimized reward signal, and it is trained collaboratively with a dual-channel neural network to gradually acquire the ability to generate dynamic tasks.

[0075] Furthermore, the steps of adjusting task metrics in real time based on the current state using a dynamic task generator include:

[0076] Step 21, record the current state as ,in, This represents the average root mean square spot radius. This represents the radius of the maximum root mean square spot. These represent the Zernike aberration coefficients for the 3rd, 5th, and 8th terms, respectively; the RMS error distribution is obtained by tracing the feature rays.

[0077] Step 22: Analyze the aberrations to obtain the Zernike aberration coefficients of the dominant aberrations;

[0078] Step 23, calculate the aberration sensitivity using the following formula:

[0079] ,

[0080] In the formula, Let k be the aberration sensitivity term. Z represents the root mean square error combining defocus, spherical aberration, and distortion. k Let k be the Zernike aberration coefficient;

[0081] Step 24, perform policy network decision-making, the calculation formula is:

[0082] ,

[0083] In the formula, Represents the action value function. Indicates an action, This represents the set of trainable parameters for a Q-network.

[0084] Step 25, calculate the reward function, the formula is:

[0085] ,

[0086] In the formula, Indicates the reward value. , Indicates the weighting coefficient. This represents the full field-of-view correction factor. Indicates the actual field of view of the current camera. Indicates the target's field of view;

[0087] Step 26, score the physical feasibility. Make a judgment, if , For physical feasibility scoring thresholds, such as setting... This involves adjusting the step size of the task indicators, reducing the difficulty of optimizing the field of view, tightening the aperture to reduce aberrations, and improving manufacturability, such as... , If the physical feasibility score is... If so, the current task indicators and optimization step size remain unchanged, and no constraint adjustments are made;

[0088] The updated task marker is as follows ,in, This indicates the updated field of view. This indicates the updated aperture number. Indicates the effective focal length.

[0089] Step 3: Input the task metrics output by the dynamic task generator into the dual-channel neural network for joint training, and obtain the optical parameters based on the trained dual-channel neural network.

[0090] Furthermore, the steps of inputting the task metrics output by the dynamic task generator into the dual-channel neural network for joint training include:

[0091] Step 31, set the current task metrics Inputting this into the performance prediction channel yields the optical system structure parameter vector, represented as follows:

[0092] ,

[0093] In the formula, This represents the performance prediction channel model; This represents a vector of structural parameters of the optical system. Indicates the first The radius of curvature of the spherical / aspherical surface. This indicates the center thickness of the lens or the air gap between lenses. Indicates the first The type of optical glass used in block lenses;

[0094] Step 32, set the optical system structural parameters The vector input is fed into the physical constraint channel to obtain the physical feasibility score, which is represented as follows:

[0095] ,

[0096] In the formula, Represents a physical constraint channel model;

[0097] Step 33, calculate the total loss, the formula is:

[0098] ,

[0099] In the formula, Indicates loss of optical performance. Represents physical constraint loss. This represents the aberration sensitivity-weighted loss. The physical constraint loss weighting coefficient, The aberration sensitivity loss weighting coefficient can be set. ;

[0100] Step 34: Update the parameters of the two-channel neural network using the AdamW optimizer. The updated formula is:

[0101] ,

[0102] In the formula, This indicates the step size for each parameter update. This represents the gradient operator of the loss function with respect to the parameter θ.

[0103] Specifically, the expression for the optical performance loss is:

[0104] ,

[0105] In the formula, N is the number of training samples. This is the vector of optical parameters predicted by the network. To supervise the true optical parameter vector of the sample, Norm operations.

[0106] Specifically, the expression for the physical constraint loss is:

[0107] ,

[0108] In the formula, It is a rectified linear activation function that penalizes only when constraints are violated; For the allowable range of structural parameters and process parameters (minimum / maximum limits); The current optical structure parameters output by the network; For optical surface gradient (change in surface slope); Let L2 be the surface gradient, representing the steepness of the surface; This represents the maximum allowable surface gradient (process upper limit).

[0109] Specifically, the expression for the aberration sensitivity weighted loss is:

[0110] ,

[0111] In the formula, Weights are optimized for each field of view / wavelength, where RMS is the root mean square spot radius. This is the sensitivity index amplification term.

[0112] Step 4: Embed lightweight ray tracing during training, extract Zernike aberration coefficients based on the tracing results as real-time feedback signals, and verify the optical performance of the dynamic task generator output.

[0113] In one example, lightweight ray tracing is performed to simulate the propagation path of light in an optical system. The field of view of the feature ray is set at the on-axis (0°), 0.7 field of view, and edge field of view, with selected wavelengths including 486nm, 587nm, and 656nm. The RMS spot radius is calculated using the following formula:

[0114] ,

[0115] In the formula, Where is the root mean square radius of the spot, and N is the total number of sampled rays. , For the first The coordinates of the point where the ray falls on the image plane. , Let be the coordinates of the centroid of the point where the light ray falls.

[0116] Based on the tracking results, the Zernike aberration coefficients are fitted using the following formula:

[0117] ,

[0118] In the formula, For the first Xiang Zernick aberration coefficient, This is the double integral over the pupil region. Let wavefront aberration function be used. For the first Xiang Zernike polynomial basis functions.

[0119] Finally, the average root mean square radius of the spot was obtained. Maximum root mean square spot radius and Zernike aberration coefficient .

[0120] Step 5: Calculate the aberration sensitivity based on the Zernike aberration coefficients from the real-time ray tracing verification feedback, prioritize the suppression of dominant aberrations based on the aberration sensitivity, and dynamically adjust the weights of the optimization strategy based on the manufacturing feasibility feedback from the physical constraint channel output.

[0121] Furthermore, the steps of dynamically adjusting the weights of the optimization strategy based on the manufacturing feasibility feedback from the physical constraint channel output include:

[0122] Step 51, analyze the current RMS distribution, if Then adjust the aberration weights using the following formula:

[0123] ,

[0124] In the formula, This represents the RMS spot radius at the edge of the field of view. This represents the RMS spot radius of the central field of view. The optimized weights for coma are represented;

[0125] Step 52, score based on physical feasibility. Adaptively adjust the learning rate of a dual-channel neural network ,like Then adjust to: , ;like Then adjust to: ;

[0126] Step 53: Adjust the weighting coefficients of the total loss function.

[0127] Based on the aberration weight adjustment results and the physical feasibility score, update the weighting coefficients of the performance term, physical constraint term, and aberration sensitivity term in the total loss function. When the physical feasibility score is high, reduce the weighting coefficient of the physical constraint loss. Increase the weighting coefficient of aberration sensitivity loss When the physical feasibility score is low, increase the weighting coefficient of the physical constraint loss. aberration sensitivity loss weighting coefficient For example, when the physical feasibility score is... When it is 0.7, Adjust from 0.5 to 0.4, The weight was adjusted from 0.2 to 0.3, completing the weight update of this round of optimization strategy.

[0128] Step 6: When the closed-loop iteration formed by dual-channel joint training, real-time ray tracing verification and dynamic optimization strategy adjustment meets the preset convergence condition, the lens design result that meets both the requirements of optical performance and manufacturing feasibility is output.

[0129] Furthermore, when the closed-loop iteration formed by dual-channel joint training, real-time ray tracing verification, and dynamic optimization strategy adjustment satisfies the preset convergence condition, the steps to output a lens design result that meets both optical performance and manufacturing feasibility requirements include:

[0130] When physical feasibility score And maximum field of view Then it is considered that the convergence condition is met. Indicates the operating wavelength (e.g.) ), calculate the tolerance sensitivity matrix , represented as:

[0131]

[0132] Based on the tolerance sensitivity matrix The cost is estimated using the following formula:

[0133] ,

[0134] In the formula, For the total cost of the lens, For the first Lens material cost Weights are used to penalize complexity. This is the design complexity function calculated based on the tolerance sensitivity matrix;

[0135] Generate a lens design report, including project design specifications, core optical system parameters, image quality simulation analysis, manufacturing cost estimates, and final design achievement conclusions and optimization suggestions.

[0136] In one example, if the RMS spot radius of the edge field of view is larger than that of the RMS spot radius of the center field of view, it indicates that the imaging quality of the lens across the entire field of view is unbalanced. Therefore, it is concluded that coma should be corrected first to improve the imaging quality of the edge field of view, and it is recommended to increase the coma optimization weight.

[0137] If physical feasibility score This indicates that the lens structure has good manufacturability, and it can be concluded that optical imaging performance can be improved first. The optimization suggestions are: keep the learning rate unchanged, reduce the weight of physical constraint loss, and increase the weight of aberration sensitivity loss.

[0138] If physical feasibility score This indicates that the lens structure is difficult to manufacture and mass-produce, leading to the conclusion that the manufacturability of the structure should be prioritized. The optimization suggestions are: reduce the learning rate, increase the weight of physical constraint loss, and reduce the weight of aberration sensitivity loss.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0140] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A lens design method based on deep learning, characterized in that, include: Construct a two-channel neural network for lens design, wherein the two-channel neural network includes a performance prediction channel and a physical constraint channel; The dynamic task generator adjusts task metrics in real time based on the current status. The task metrics output by the dynamic task generator are input into the dual-channel neural network for joint training, and the optical parameters are obtained based on the trained dual-channel neural network. Lightweight ray tracing is embedded during training, and Zernike aberration coefficients are extracted based on the tracing results as real-time feedback signals to verify the optical performance of the dynamic task generator output. Aberration sensitivity is calculated based on Zernike aberration coefficients from real-time ray tracing verification feedback. Dominant aberrations are suppressed first based on aberration sensitivity, and the weights of the optimization strategy are dynamically adjusted based on manufacturing feasibility feedback from the physical constraint channel output. When the closed-loop iteration formed by dual-channel joint training, real-time ray tracing verification, and dynamic optimization strategy adjustment meets the preset convergence conditions, the output lens design result satisfies both optical performance and manufacturing feasibility requirements.

2. The lens design method based on deep learning according to claim 1, characterized in that, The steps involved in using a dynamic task generator to adjust task metrics in real time based on the current status include: Step 21, record the current state as ,in, This represents the average root mean square spot radius. This represents the radius of the maximum root mean square spot. These represent the Zernike aberration coefficients for terms 3, 5, and 8, respectively; the RMS error distribution is obtained by tracing the feature rays. Step 22: Analyze the aberrations to obtain the Zernike aberration coefficients of the dominant aberrations; Step 23, calculate aberration sensitivity; Step 24, perform policy network decision-making, the calculation formula is: ， In the formula, Represents the action value function. Indicates an action, This represents the set of trainable parameters for a Q-network. Step 25, calculate the reward function, the formula is: ， In the formula, Indicates the reward value. , Indicates the weighting coefficient. This represents the full field-of-view correction factor. Indicates the actual field of view of the current camera. Indicates the target's field of view; Step 26, score the physical feasibility. Make a judgment, if , If the physical feasibility score threshold is used, then the step size of the task indicators is adjusted to make... , ; The updated task marker is as follows ,in, This indicates the updated field of view. This indicates the updated aperture number. Indicates the effective focal length.

3. The deep learning-based lens design method according to claim 1, characterized in that, The steps for jointly training a dual-channel neural network by inputting the task metrics output by the dynamic task generator include: Step 31: Input the current task metrics into the performance prediction channel to obtain the optical system structure parameter vector, represented as: ， In the formula, This represents the performance prediction channel model; This represents a vector of structural parameters of the optical system. Indicates the first The radius of curvature of the spherical / aspherical surface. This indicates the center thickness of the lens or the air gap between lenses. Indicates the first The type of optical glass used in block lenses; Step 32: Input the optical system structural parameter vector into the physical constraint channel to obtain the physical feasibility score, expressed as: ， In the formula, Represents a physical constraint channel model; Step 33, calculate the total loss, the formula is: ， In the formula, Indicates loss of optical performance. Indicates physical constraint loss, This represents the aberration sensitivity-weighted loss. The physical constraint loss weighting coefficient, This is the weighting coefficient for aberration sensitivity loss; Step 34: Update the parameters of the two-channel neural network using the AdamW optimizer. The updated formula is: ， In the formula, This indicates the step size for each parameter update. This represents the gradient operator of the loss function with respect to the parameter θ.

4. The lens design method based on deep learning according to claim 1, characterized in that, The steps for dynamically adjusting the weights of the optimization strategy based on the manufacturing feasibility feedback from the physical constraint channel output include: Step 51, analyze the current RMS distribution, if Then adjust the aberration weights using the following formula: ， In the formula, This represents the RMS spot radius at the edge of the field of view. This represents the RMS spot radius of the central field of view. The optimized weights for coma are represented; Step 52, score based on physical feasibility. Adaptive learning rate adjustment ,like Then adjust to: , ;like Then adjust to: ; Step 53: Adjust the weighting coefficients of the total loss function.

5. The deep learning-based lens design method according to claim 1, characterized in that, When the closed-loop iteration formed by dual-channel joint training, real-time ray tracing verification, and dynamic optimization strategy adjustment meets the preset convergence condition, the steps to output a lens design result that meets both optical performance and manufacturing feasibility requirements include: When physical feasibility score And maximum field of view Then it is considered that the convergence condition is met. Indicate the operating wavelength and calculate the tolerance sensitivity matrix. , is represented as: ， Based on the tolerance sensitivity matrix The cost is estimated using the following formula: ， In the formula, For the total cost of the lens, For the first Lens material cost Weights are used to penalize complexity. This is the design complexity function calculated based on the tolerance sensitivity matrix; Generate a lens design report, including project design specifications, core optical system parameters, image quality simulation analysis, manufacturing cost estimates, and final design achievement conclusions and optimization suggestions.

Images