A method and apparatus for removing backfetch error in aspheric measurements based on graph-to-graph neural networks.

Abstract

A method and apparatus for removing backhaul errors in aspheric surface measurement based on graph-to-graph neural networks include: building a virtual interferometric system, creating a graph-to-graph backhaul error removal dataset, building a graph-to-graph neural network model, training the graph-to-graph neural network model, and solving for the actual aspheric surface shape error. This invention utilizes computer simulation of an optical system using the non-zero-position compensation method to generate a dataset, completing the training of the graph-to-graph neural network. This is convenient and fast. For a set of interference wavefront data from an actual optical system, backhaul errors can be eliminated without any prior knowledge or preprocessing, demonstrating good universality. Using a graph-to-graph neural network to remove backhaul errors preserves high-frequency information in the interference wavefront image and the distribution of the measured aspheric surface shape, resulting in higher detection accuracy. In the solution process, there are no inherent system errors, and the solution accuracy does not depend on the modeling accuracy of the system, avoiding complex assembly and calibration processes, and achieving fast and high-precision measurement of aspheric surface shape errors.

Description

Technical Field

[0001] This invention relates to the field of optical measurement technology, and in particular to a method for removing backlash error in aspheric measurement based on graph-to-graph neural networks, and a device for removing backlash error in aspheric measurement based on graph-to-graph neural networks, which is mainly used for rapid and high-precision measurement of aspheric surface shape error. Background Technology

[0002] An aspherical mirror is an optical element whose surface shape is not a typical sphere, but a complex curve designed to meet specific requirements. Unlike spherical mirrors, the curvature of an aspherical mirror varies at different locations. This aspherical design gives it superior optical performance, effectively reducing optical aberrations such as spherical aberration and astigmatism, thus improving image quality. It is widely used in complex and precise optical systems such as high-end camera lenses, astronomical telescopes, and medical imaging equipment. Therefore, the precise measurement of aspherical surface shape errors places stringent requirements.

[0003] Among various measurement techniques, interferometry offers advantages such as high precision, non-contact operation, and short measurement time. Non-zero compensation in interferometry refers to the inability of the compensator to completely compensate for aberrations caused by an ideal aspherical surface, resulting in residual wavefront aberration. This causes the measurement light to fail to return along its original path. Thus, a one-to-one correspondence between the compensator and the measured mirror is not required, reducing the design and adjustment requirements for the compensator. Non-zero compensation is currently the main method for expanding dynamic range and achieving high-precision measurement. However, the design of residual errors introduces backlash error into the optical system, which is extremely detrimental to precision measurement when the aspherical surface shape error is large.

[0004] To eliminate the influence of backlash error, researchers have proposed two methods in aspheric measurement using non-zero-position compensation interferometry: the half-method based on ray tracing and the inverse optimization method. When the aspheric surface shape error is large, the half-method itself has a significant systematic error. The accuracy of the inverse optimization method in solving the aspheric surface shape error depends heavily on the accuracy of the system modeling. Factors such as inaccuracies in optical component parameters and inconsistencies between component positions and the actual system, the randomness of temperature, humidity, and airflow velocity, and the selection of detector resolution and the number of Zernike polynomial terms in the inverse optimization algorithm all affect the accuracy of the final measurement results. Furthermore, the inverse optimization method requires multiple iterative optimizations of the optimization variables, significantly reducing the efficiency of solving the surface shape error and making it unsuitable for rapid measurements in industrial production.

[0005] Chinese patent application No. 202010105329X presents a technical solution proposed by the inventor: a method for eliminating backlash error in high-precision aspherical surface shape error measurement. It utilizes computer simulation to generate several sets of surface shape error data for the measured surface and interference wavefront data of the optical system. These data are one-dimensional matrices. A neural network for eliminating backlash error is constructed, trained, and then used to solve for the aspherical surface shape error, yielding the one-dimensional matrix. While this fully connected neural network can directly obtain the measurement result by inputting the Zernike polynomial coefficients of the interference wavefront and outputting the Zernike coefficients of the measured surface shape, its accuracy is limited by the finite number of Zernike polynomial coefficients, resulting in the loss of certain high-frequency information and reducing the accuracy of the measurement results. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a method for removing backlash error in aspheric measurement based on graph-to-graph neural networks. This method has good universality, can eliminate backlash error, and when combined with the ray tracing function of optical software, reduces the loss of high-frequency information, thereby achieving fast and high-precision measurement of aspheric surface shape error.

[0007] The technical solution of this invention is: a method for removing backhaul errors in aspheric measurements based on graph-to-graph neural networks, which includes the following steps:

[0008] (1) Build a virtual interference system. Based on the interference optical path used in the actual aspherical surface shape error detection work, build a virtual interference system in the optical tracking software.

[0009] (2) Create a graph-to-graph back-pass error elimination dataset. Apply a random surface error distribution to the test piece in the optical tracking software, calculate the corresponding interference wavefront, save the two-dimensional information calculated by the simulation software, repeat the simulation a large number of times, and create a graph-to-graph back-pass error elimination dataset.

[0010] (3) Build a graph-to-graph neural network model. Use the interferometric wavefront image as the network data and the distribution of the measured aspherical surface shape as the network output. Make the input dimension of the network the same as the dimension of the interferometric wavefront image, and make the output dimension of the network the same as the dimension of the surface shape error distribution.

[0011] (4) Train the graph-to-graph neural network model. Train the graph-to-graph neural network on the graph-to-graph backhaul error elimination dataset to obtain the graph-to-graph backhaul error elimination network model.

[0012] (5) Solve for the actual aspheric surface shape error. Put the interference wavefront image obtained in the actual detection work into the trained graph-to-graph backhaul error elimination network model to obtain the surface shape distribution result of the actual aspheric surface after eliminating the backhaul error.

[0013] This invention establishes a method and apparatus for removing backhaul errors in non-zero interferometric aspherical measurements based on graph-to-graph neural networks by sequentially constructing a virtual interferometric system, creating a graph-to-graph backhaul error removal dataset, constructing a graph-to-graph neural network model, training the graph-to-graph neural network model, and solving for actual aspherical surface shape errors. This invention utilizes computer simulation of an optical system using the non-zero compensation method to generate a dataset and completes the training of a graph-to-graph neural network. This is convenient and fast. For a set of interference wavefront data from a real optical system, no prior knowledge or preprocessing is required to eliminate backlash errors in the actual optical system, demonstrating good universality. This invention uses a graph-to-graph neural network to eliminate backlash errors, preserving high-frequency information in the interference wavefront image and the distribution of the measured aspherical surface shape, resulting in higher detection accuracy. In the process of solving for aspherical surface shape errors, there are no inherent system errors, and the solution accuracy does not depend on the modeling accuracy of the system. Inaccuracies in optical component parameters and inconsistencies between component positions and the actual system, as well as the randomness of temperature, humidity, and airflow velocity, do not affect the final result. This avoids complex assembly and calibration processes, achieving fast and high-precision measurement of aspherical surface shape errors.

[0014] A backflash error removal device for aspherical measurement based on graph-to-graph neural network is also provided, which includes:

[0015] A virtual interferometric system module is built, configured to build a virtual interferometric system in optical tracing software based on the interferometric optical path used in the actual aspherical surface shape error detection work; a graph-to-graph backhaul error elimination dataset module is created, configured to apply a random surface shape error distribution to the test piece in optical tracing software, calculate the corresponding interference wavefront image, save the two-dimensional information calculated by the simulation software, repeat the simulation a large number of times, and create a graph-to-graph backhaul error elimination dataset;

[0016] The module for building a graph-to-graph neural network model is configured to build a graph-to-graph neural network model. The interferometric wavefront image is used as the network data, and the distribution of the measured aspherical surface shape is used as the network output. The input dimension of the network is the same as the dimension of the interferometric wavefront image, and the output dimension of the network is the same as the dimension of the surface shape error distribution.

[0017] The training graph-to-graph neural network model module is configured to train the graph-to-graph neural network on the graph-to-graph backhaul error elimination dataset to obtain the graph-to-graph backhaul error elimination network model.

[0018] The module for solving the actual aspherical surface shape error is configured to input the interference wavefront image obtained in the actual detection work into the trained graph-to-graph backhaul error removal network model, so as to obtain the surface shape distribution result of the actual aspherical surface after removing the backhaul error. Attached Figure Description

[0019] Figure 1 This is a flowchart of the backhaul error removal method for aspheric measurement based on graph-to-graph neural networks according to the present invention.

[0020] Figure 2 It is the interference wavefront diagram obtained from the measured aspherical surface in interferometry.

[0021] Figure 3 It is the true value of the surface shape distribution of the aspherical surface being measured.

[0022] Figure 4 It is the result of the graph-to-graph neural network outputting the calculated surface shape error of the measured aspherical surface.

[0023] Figure 5 It represents the difference between the aspherical surface shape error result output by the graph-to-graph neural network and the true value. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of 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 skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] It should be noted that the term "comprising" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products or devices.

[0026] like Figure 1 As shown, this method for removing backhaul errors in aspheric measurements based on graph-to-graph neural networks includes the following steps:

[0027] (1) Build a virtual interference system. Based on the interference optical path used in the actual aspherical surface shape error detection work, build a virtual interference system in the optical tracking software.

[0028] (2) Create a graph-to-graph back-pass error elimination dataset. Apply a random surface error distribution to the test piece in the optical tracking software, calculate the corresponding interference wavefront, save the two-dimensional information calculated by the simulation software, repeat the simulation a large number of times, and create a graph-to-graph back-pass error elimination dataset.

[0029] (3) Build a graph-to-graph neural network model. Use the interferometric wavefront image as the network data and the distribution of the measured aspherical surface shape as the network output. Make the input dimension of the network the same as the dimension of the interferometric wavefront image, and make the output dimension of the network the same as the dimension of the surface shape error distribution.

[0030] (4) Train the graph-to-graph neural network model. Train the graph-to-graph neural network on the graph-to-graph backhaul error elimination dataset to obtain the graph-to-graph backhaul error elimination network model.

[0031] (5) Solve for the actual aspheric surface shape error. Put the interference wavefront image obtained in the actual detection work into the trained graph-to-graph backhaul error elimination network model to obtain the surface shape distribution result of the actual aspheric surface after eliminating the backhaul error.

[0032] This invention establishes a method and apparatus for removing backhaul errors in non-zero interferometric aspherical measurements based on graph-to-graph neural networks by sequentially constructing a virtual interferometric system, creating a graph-to-graph backhaul error removal dataset, constructing a graph-to-graph neural network model, training the graph-to-graph neural network model, and solving for actual aspherical surface shape errors. This invention utilizes computer simulation of an optical system using the non-zero compensation method to generate a dataset and completes the training of a graph-to-graph neural network. This is convenient and fast. For a set of interference wavefront data from a real optical system, no prior knowledge or preprocessing is required to eliminate backlash errors in the actual optical system, demonstrating good universality. This invention uses a graph-to-graph neural network to eliminate backlash errors, preserving high-frequency information in the interference wavefront image and the distribution of the measured aspherical surface shape, resulting in higher detection accuracy. In the process of solving for aspherical surface shape errors, there are no inherent system errors, and the solution accuracy does not depend on the modeling accuracy of the system. Inaccuracies in optical component parameters and inconsistencies between component positions and the actual system, as well as the randomness of temperature, humidity, and airflow velocity, do not affect the final result. This avoids complex assembly and calibration processes, achieving fast and high-precision measurement of aspherical surface shape errors.

[0033] Preferably, in step (1), the virtual interferometer built by the optical tracing software is consistent with the actual detection optical path.

[0034] Preferably, in step (2), random surface shape errors are added to the aspheric surface under test in the virtual interferometer by adding Zernike surfaces, and the corresponding interference wavefront image is obtained after running ray tracing. The resolution of the interference wavefront image and the surface shape error of the aspheric surface under test is above 224×224.

[0035] Preferably, in step (3), the graph-to-graph neural network model is the U-Net image generation network. This ensures that the network has high complexity and performance.

[0036] Preferably, in step (4), the training of the graph-to-graph neural network involves constantly monitoring changes in the loss function and accuracy to prevent the network model from being undertrained or overfitted.

[0037] Preferably, in step (5), the actual detected interference wavefront is scaled to the same size as the dataset so that it can be put into the network to correctly solve the surface error.

[0038] Preferably, in step (1), a Fizzo interferometer consistent with the actual optical path is constructed using Zemax software.

[0039] Preferably, in step (2), 78 Zernike coefficients are randomly generated, and the surface shape PV generated by these coefficients is between 2λ and 0.05λ. In the Zemax software, the randomly generated Zernike polynomial coefficients are substituted into the aspheric surface under test in the form of Zernike sag surfaces, and the surface shape generated by this set of coefficients is used as the true value of the surface shape error for later use. The optical path tracing is run to record the interference wavefront image. The true value of the aspheric surface shape error and the interference wavefront image are used as a set of data with a resolution of 224×224. The experiment is repeated to obtain 6052 sets of data, which together constitute the backhaul error elimination dataset. In step (3), a swin-unet network model is built, with both input and output dimensions of 224x224.

[0040] Preferably, in step (4), the swin-unet network model is trained on 6052 sets of data, the batch size is set to 32, a total of 300 training rounds are conducted, the optimizer is Adam, and the learning rate is 1e-4.

[0041] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When executed, the program includes the steps of the methods of the above embodiments. The storage medium can be ROM / RAM, magnetic disk, optical disk, memory card, etc. Therefore, corresponding to the method of the present invention, the present invention also includes a backhaul error removal device for aspherical measurement based on graph-to-graph neural networks. This device is typically represented in the form of functional modules corresponding to the steps of the method. The device includes:

[0042] A virtual interferometric system module is built, configured to build a virtual interferometric system in optical tracing software based on the interferometric optical path used in the actual aspherical surface shape error detection work; a graph-to-graph backhaul error elimination dataset module is created, configured to apply a random surface shape error distribution to the test piece in optical tracing software, calculate the corresponding interference wavefront image, save the two-dimensional information calculated by the simulation software, repeat the simulation a large number of times, and create a graph-to-graph backhaul error elimination dataset;

[0043] The module for building a graph-to-graph neural network model is configured to build a graph-to-graph neural network model. The interferometric wavefront image is used as the network data, and the distribution of the measured aspherical surface shape is used as the network output. The input dimension of the network is the same as the dimension of the interferometric wavefront image, and the output dimension of the network is the same as the dimension of the surface shape error distribution.

[0044] The training graph-to-graph neural network model module is configured to train the graph-to-graph neural network on the graph-to-graph backhaul error elimination dataset to obtain the graph-to-graph backhaul error elimination network model.

[0045] The module for solving the actual aspherical surface shape error is configured to input the interference wavefront image obtained in the actual detection work into the trained graph-to-graph backhaul error removal network model, so as to obtain the surface shape distribution result of the actual aspherical surface after removing the backhaul error.

[0046] The following details a specific embodiment of the present invention. A method for removing backhaul errors in non-zero interferometric aspherical measurements based on graph-to-graph neural networks is implemented as follows:

[0047] The process for removing backflash errors in nonzero interferometric aspherical measurements based on graph-to-graph neural networks is as follows: Figure 1 As shown, the specific implementation steps are as follows:

[0048] Step 1: Building a Virtual Interference System

[0049] In this example, a Fizeau interferometer that matches the actual optical path was built using Zemax software.

[0050] Step 2: Create a dataset for removing retracement errors

[0051] In this example, 78 Zernike coefficients are randomly generated, ensuring that the resulting surface shape (PV) is between 2λ and 0.05λ. In Zemax, the randomly generated Zernike polynomial coefficients are substituted into the measured aspherical surface in the form of a Zernike sag surface, and the surface shape generated by this set of coefficients is used as the true value for the surface shape error. Figure 3 As shown. Run optical path tracing and record the interference wavefront diagram, as shown. Figure 2 As shown, the true values ​​of the aspherical surface shape error and the interference wavefront diagram are used as a set of data with a resolution of 224*224. The experiment was repeated to obtain 6052 sets of data, which together constitute the backflash error elimination dataset.

[0052] Step 3: Building a graph-to-graph neural network model

[0053] In this example, a swin-unet network model is built, with both input and output dimensions of 224*224.

[0054] Step 4: Training the graph-to-graph neural network model

[0055] In this example, the swin-unet network model was trained on 6052 datasets with a batch size of 32 for a total of 300 training epochs. The optimizer was Adam, and the learning rate was 1e-4.

[0056] Step 5: Solving for the actual aspherical surface shape error

[0057] In the example, data from the validation set is fed into the network model, and the resulting surface error distribution is as follows: Figure 4 As shown, its error compared to the true value is as follows: Figure 5 As shown, the backlash error removal of the aspherical surface shape error distribution has been completed.

[0058] The beneficial effects of the present invention are as follows:

[0059] 1. This invention uses a graph neural network to remove backlash errors in aspherical surface shape error detection. Without Zernike fitting, it directly correlates the distribution of aspherical surface shape errors with the interference wavefront diagram, effectively preserving the high-frequency information.

[0060] 2. The graph-to-graph network used in this invention, once trained, can eliminate backlash errors in a set of actual optical system interference wavefront data without any prior knowledge or preprocessing, thus enabling rapid measurement of aspherical surface shape errors.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for removing backfetch error in aspherical measurement based on graph-to-graph neural networks, characterized in that: It includes the following steps: (1) Build a virtual interference system. Based on the interference optical path used in the actual aspherical surface shape error detection work, build a virtual interference system in the optical tracking software. (2) Create a graph-to-graph back-pass error elimination dataset. Apply a random surface error distribution to the test piece in the optical tracking software, calculate the corresponding interference wavefront, save the two-dimensional information calculated by the simulation software, repeat the simulation a large number of times, and create a graph-to-graph back-pass error elimination dataset. (3) Build a graph-to-graph neural network model. Use the interferometric wavefront image as the network data and the distribution of the measured aspherical surface shape as the network output. Make the input dimension of the network the same as the dimension of the interferometric wavefront image, and make the output dimension of the network the same as the dimension of the surface shape error distribution. (4) Train the graph-to-graph neural network model. Train the graph-to-graph neural network on the graph-to-graph backhaul error elimination dataset to obtain the graph-to-graph backhaul error elimination network model. (5) Solve for the actual aspheric surface shape error. Put the interference wavefront image obtained in the actual detection work into the trained graph-to-graph backhaul error elimination network model to obtain the surface shape distribution result of the actual aspheric surface after eliminating the backhaul error.

2. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 1, characterized in that: In step (1), the virtual interferometer built by the optical tracing software is consistent with the actual detection optical path.

3. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 2, characterized in that: In step (2), random surface shape errors are added to the aspheric surface under test in the virtual interferometer by adding Zernike surfaces. After running ray tracing, the corresponding interference wavefront image is obtained. The resolution of the interference wavefront image and the surface shape error of the aspheric surface under test is above 224×224.

4. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 2, characterized in that: In step (3), the graph-to-graph neural network model is the U-Net image generation network.

5. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 3 or 4, characterized in that: In step (4), the training of the graph-to-graph neural network involves constantly monitoring the changes in the loss function and accuracy to prevent the network model from being undertrained or overfitted.

6. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 5, characterized in that: In step (5), the actual detected interference wavefront is scaled to the same size as the dataset so that it can be put into the network to correctly solve the surface error.

7. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 1, characterized in that: In step (1), a Fizzo-type interferometer that matches the actual optical path is constructed using Zemax software.

8. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 7, characterized in that: In step (2), 78 Zernike coefficients are randomly generated, and the surface shape PV generated by these coefficients is between 2λ and 0.05λ. In the Zemax software, the randomly generated Zernike polynomial coefficients are substituted into the aspheric surface under test in the form of Zernike sag surfaces, and the surface shape generated by the Zernike polynomial coefficients is used as the true value of the surface shape error for later use. The optical path tracing is run to record the interference wavefront image. The true value of the aspheric surface shape error and the interference wavefront image are used as a set of data with a resolution of 224x224. The experiment is repeated to obtain 6052 sets of data, which together constitute the backhaul error elimination dataset. In step (3), a swin-unet network model is built with input and output dimensions of 224x224.

9. The method for removing backhaul error in aspheric measurement based on graph-to-graph neural networks according to claim 8, characterized in that: In step (4), the swin-unet network model is trained on 6052 sets of data, the batch size is set to 32, a total of 300 training rounds are conducted, the optimizer is Adam, and the learning rate is 1e-4.

10. A backfetch error removal device for aspherical measurement based on graph-to-graph neural network, characterized in that: It includes: A virtual interferometric system module is built, configured to build a virtual interferometric system in optical tracing software based on the interferometric optical path used in the actual aspherical surface shape error detection work; a graph-to-graph backhaul error elimination dataset module is created, configured to apply a random surface shape error distribution to the test piece in optical tracing software, calculate the corresponding interference wavefront image, save the two-dimensional information calculated by the simulation software, repeat the simulation a large number of times, and create a graph-to-graph backhaul error elimination dataset; The module for building a graph-to-graph neural network model is configured to build a graph-to-graph neural network model. The interferometric wavefront image is used as the network data, and the distribution of the measured aspherical surface shape is used as the network output. The input dimension of the network is the same as the dimension of the interferometric wavefront image, and the output dimension of the network is the same as the dimension of the surface shape error distribution. The training graph-to-graph neural network model module is configured to train the graph-to-graph neural network on the graph-to-graph backhaul error elimination dataset to obtain the graph-to-graph backhaul error elimination network model. The module for solving the actual aspherical surface shape error is configured to input the interference wavefront image obtained in the actual detection work into the trained graph-to-graph backhaul error removal network model, so as to obtain the surface shape distribution result of the actual aspherical surface after removing the backhaul error.

Images