Method for determining polishing process parameters based on image quality indexes based on neural network

Abstract

The present invention relates to the field of optical processing technology, and in particular to a method for determining polishing process parameters based on image quality indicators based on a neural network. The method comprises: determining the structural parameters of an optical system according to specific requirements; marking optical components requiring process decisions in the optical system and recording historical processing data; constructing an optical system, placing the marked optical components in the optical system, debugging to obtain actual image quality indicators, and collecting the degree of deterioration compared to the ideal system image quality; repeating the above steps to obtain a data set; using the image quality deterioration data set as input and the component process parameters as output, inputting the data into a modular convolutional neural network for training to obtain a predictive network model; and when a manufacturing task for an optical system component arrives, inputting the system type and image quality deterioration tolerance into the predictive network model to obtain polishing process parameters that meet the image quality deterioration tolerance. The advantages of the method are: rapid decision-making, low computational time, and high ease of understanding.

Description

Technical Field

[0001] The present invention relates to the field of optical processing technology, and in particular to a method for determining polishing process parameters based on image quality indicators of a neural network. Background Art

[0002] In any manufacturing process, residual manufacturing errors are inevitable on the surface of components. Installing optical components with these errors in a system affects the imaging quality of the optical system. A paper published in the journal Photonics, "Analysis and Prediction of Image Quality Degradation Caused by Diffraction of Infrared Optical System Turning Marks," explains the impact of manufacturing errors generated by single-point diamond turning on diffraction. This deterministic process produces manufacturing errors with a periodicity on the micron scale, which can produce a significant diffraction effect on visible light. This can be explained using diffraction theory, which infers which manufacturing error indicators are suitable by reducing diffraction intensity, and thus selects process parameters.

[0003] The paper "Surface variation analysis of freeform optical systems over surface frequency bands for prescribed wavefront errors," published in the journal Optics & Laser Technology, explains the impact of annular-like manufacturing errors on optical components and uses geometric optics to deduce the range within which the manufacturing error level should be.

[0004] The characteristic period of manufacturing errors generated by magnetorheological process is in the millimeter level, and the characteristic period of manufacturing errors generated by small grinding head process is in the centimeter level. Moreover, most of the manufacturing error characteristics generated by this process are grating-like errors, and their period scale is much larger than the wavelength. The diffraction theory cannot produce obvious diffraction effects. In the absence of a model to accurately describe the suitability of the process, blind box fitting tools such as neural networks are relatively suitable. Therefore, a method for determining polishing process parameters based on image quality indicators based on neural networks should be developed. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a method for determining polishing process parameters based on image quality indicators of a neural network.

[0006] The present invention aims to provide a method for determining polishing process parameters based on image quality indicators of a neural network, which specifically includes the following steps:

[0007] S1. Determine the structural parameters of the optical system according to specific needs;

[0008] S2. Marking the optical element in the optical system that requires process decision-making and recording the historical processing data of the optical element;

[0009] S3. Build the optical system, place the marked optical element in step S2 in the optical system, debug to obtain the actual image quality index, and collect data on the degree of deterioration compared with the ideal system image quality;

[0010] S4 repeats steps S2 ~ S3 to obtain a data set of process parameters of the optical element and a data set of image quality deterioration;

[0011] S5. Using the image quality deterioration dataset as the input dataset and the component process parameters as the output dataset, the dataset is input into a modular convolutional neural network for training to obtain a network model for predicting optical component polishing process parameters;

[0012] S6. According to the requirements of the process decision task, the image quality deterioration tolerance of the optical element is obtained, and the maximum value of the image quality deterioration tolerance is input as the image quality deterioration data into the optical element polishing process parameter prediction network model to obtain the polishing process parameters that meet the image quality deterioration tolerance.

[0013] Preferably, the optical system structural parameters include the type of the optical system, and the relative positions, placement angles, component surface shapes, and substrate materials of all optical components in the optical system.

[0014] Preferably, the historical processing data in step S2 includes process parameters, environmental parameters, and manufacturing error data corresponding to the process parameters;

[0015] The process parameter data include machining gauge h, removal function RF, machining angle theta, and machining dwell time dt.

[0016] Preferably, in step S4, steps S2 to S3 are repeated for at least N groups, where N≥50.

[0017] Preferably, the image quality deterioration degree dataset is used as the input dataset, and the process parameter dataset is used as the output dataset;

[0018] The modular convolutional neural network includes a backbone convolution layer and a dynamic weight layer; the backbone convolution layer is used to extract the features of image quality deterioration data; the dynamic weight layer dynamically loads different weights according to the optical system category, and is used to map the features to the process parameter space.

[0019] Preferably, the training optimization process in step S5 is as follows:

[0020] S501. Perform weighted search based on the optical system feature vector. The weight index k is determined by the following formula:

[0021]

[0022] Where x is the optical system eigenvector; The function is used to find the input parameter value that makes the function reach the maximum value, or return the index of the maximum value element in the data set;

[0023] S502. Input the image quality deterioration data into the modular convolutional neural network; the backbone convolution layer extracts the feature F of the image quality deterioration data:

[0024]

[0025] Among them, F is the image quality deterioration data feature extracted by the backbone convolution layer, and is also the input of the dynamic weight layer; Conv represents the convolution function;

[0026] The feature F of the image quality deterioration data output by the backbone convolution layer is flattened into a vector f, and the dynamically loaded weights PT k Map to the process parameter space and calculate the predicted value of the process parameter; the calculation formula is as follows:

[0027]

[0028] in, is the ReLU activation function, Bias term, is the predicted value of process parameters;

[0029] S503. Use the difference between the predicted process parameters and the actual process parameters as the loss function, expressed as follows:

[0030]

[0031] Where, h i Indicates the actual processing gauge, represents the predicted processing trajectory, Indicates the actual processing angle, represents the predicted processing angle, … represents other process parameters.

[0032] Preferably, in the training optimization process of step S5, the Adam algorithm is used as the optimization algorithm to fit the relationship between the input data set and the output data set under different optical systems, and the weight files trained by different optical systems are numbered and recorded as PT k , ; where k is the number of the optical system and W is the number of types of optical systems.

[0033] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0034] The present invention provides a method for determining polishing process parameters based on image quality indicators of a neural network, that is, a method for determining optical processing process parameters associated with imaging indicators of an optical system. Although the method of the present invention requires a large amount of data to be collected in the early stage, once the neural network calculation model is built, the acquisition of decision-making process parameters is very fast and objective. The method of determining deterministic polishing process parameters based on image quality indicators of a neural network proposed in the present invention has obvious advantages in terms of calculation time and ease of understanding. Traditional debugging determines process parameters by continuously adjusting parameters through experiments. Compared with traditional debugging methods, the process development efficiency of the present invention can be increased by 5-6 times, and is particularly suitable for precision processing scenarios of complex optical components such as aspheric surfaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 The present invention provides a flowchart of a method for determining polishing process parameters based on image quality indicators using a neural network according to an embodiment of the present invention.

[0036] Figure 2 is a removal function diagram provided according to an embodiment of the present invention.

[0037] Figure 3 is a simulated tool mark error provided according to an embodiment of the present invention.

[0038] Figure 4 This is a flowchart of modular neural network weight selection according to an embodiment of the present invention.

[0039] Figure 5 It is a flowchart of the operation of a modular neural network provided according to an embodiment of the present invention.

[0040] Figure 6 is an optical system image quality evaluation result provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, identical modules are denoted by identical reference numerals. In the case of identical reference numerals, their names and functions are also identical. Therefore, their detailed description will not be repeated.

[0042] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0043] The present invention provides a method for determining polishing process parameters based on image quality indicators of a neural network (flow chart as shown in FIG. Figure 1 ), which aims to efficiently and accurately determine the polishing process parameters of optical components to meet the strict image quality requirements of the optical system; specifically, it includes the following steps:

[0044] S1. Obtaining basic system information: Determine the structure of the optical system and the structural parameters of the optical system to be established based on specific needs;

[0045] In a specific embodiment, the optical system structural parameters include the type of optical system (such as off-axis three-mirror type, Cassegrain type, etc.), as well as the relative position, placement angle, surface shape, substrate material, etc. of all optical elements in the optical system; the above information will provide an important reference basis for subsequent process parameter decisions, ensuring that the constructed neural network can accurately reflect the actual characteristics of the optical system.

[0046] S2. Marking the optical element in the optical system that requires process decision-making and recording the historical processing data of the optical element;

[0047] The historical processing data includes process parameters and manufacturing errors corresponding to the process parameters;

[0048] Process parameter data include machining track gauge h, removal function RF, machining angle theta, machining dwell time dt, etc.; removal function such as Figure 2 As shown; for the convenience of demonstration, a magnetorheological polishing feature manufacturing error is shown in Figure 3 .

[0049] S3. Build the optical system, place the optical components required for decision-making in the system, debug to obtain actual image quality indicators, and collect data on the degree of degradation compared to the ideal system image quality;

[0050] The degree of image quality deterioration is a key indicator to measure the gap in imaging quality between actual optical systems and ideal optical systems. By collecting this data, clear input targets can be provided for the neural network, enabling it to learn how to adjust process parameters according to image quality requirements.

[0051] S4. Repeat steps S2 to S3 for at least N groups (N ≥ 50) to obtain a process parameter dataset and an image quality degradation dataset for the optical element; use the image quality degradation dataset as the input dataset and the process parameter dataset as the output dataset;

[0052] Through iterative verification of N≥50 sets of experimental data, the generalization ability of the neural network in the high-dimensional space of process parameters is ensured;

[0053] S5. Input the image quality deterioration dataset into a modular convolutional neural network (MNN) for training to obtain a network model for predicting optical component polishing process parameters. A modular convolutional neural network is a modular network that can select different weight data for network prediction based on different optical system structural parameters. Specifically, during the training phase, input and output datasets obtained under different optical system structures are trained, and the trained network weights are saved. In the next prediction process, the network structure does not need to be adjusted; only the characteristic vector of the optical system structure needs to be input, thereby achieving modular application for different optical systems.

[0054] The modular convolutional neural network consists of a backbone convolutional layer and a dynamic weight layer. The backbone convolutional layer, also called a shared convolutional layer, is the same for different types of optical systems and is used to extract the feature F of the image quality degradation data J:

[0055]

[0056] Among them, F is the image quality deterioration data feature extracted by the backbone convolution layer, and is also the input of the dynamic weight layer; Conv represents the convolution function;

[0057] The dynamic weight layer dynamically loads different weights according to the optical system category to map the features to the process parameter space. Specifically, the output F of the backbone convolution layer is flattened into a vector f, and the dynamically loaded weights are used to map the features to the process parameter space. PT k Mapping to process parameter space:

[0058]

[0059] in, is the activation function (ReLU function), Bias term, is the predicted value of the process parameters.

[0060] The training optimization process is as follows:

[0061] S501. Perform weighted search based on the optical system feature vector. The weight index k is determined by the following formula:

[0062]

[0063] Where x is the optical system eigenvector, It is a commonly used function in mathematics and computer science. Its core function is to find the input parameter value that makes a function reach the maximum value, or return the index of the maximum value element in a data set.

[0064] S502. Input the image quality deterioration data into the modular convolutional neural network; the backbone convolution layer extracts the feature F of the image quality deterioration data:

[0065]

[0066] Among them, F is the image quality deterioration data feature extracted by the backbone convolution layer, and is also the input of the subsequent dynamic weight layer. Conv represents the convolution function;

[0067] The feature F of the image quality deterioration data output by the backbone convolution layer is flattened into a vector f, and the dynamically loaded weights PT k Map to the process parameter space and calculate the predicted value of the process parameter; the calculation formula is as follows:

[0068]

[0069] in, is the activation function (ReLU function), Bias term, is the predicted value of process parameters;

[0070] S503. Use the difference between the predicted process parameters and the actual process parameters as the loss function, expressed as follows:

[0071]

[0072] Where, h i Indicates the actual processing gauge, represents the predicted processing trajectory, Indicates the actual processing angle, represents the predicted processing angle, ... represents other process parameters, which will not be introduced one by one;

[0073] The Adam algorithm is used as the optimization algorithm to fit the relationship between the input data set and the output data set under different optical systems. The weight files trained by different optical systems are numbered and recorded as PT k , ; Where k is the number of the optical system, W is the number of types of optical systems;

[0074] One-hot encoding is used to convert the optical system category into a computable vector form. For different optical systems, in a specific embodiment, taking Cassegrain and off-axis three-mirror systems as examples, one-hot encoding is used to process them to obtain the optical system category features in vector form, such as [1,0] represents the off-axis three-mirror system and [0,1] represents the Cassegrain system. The encoded optical system feature vector is matched with the corresponding module neural network weight file to ensure that the network can distinguish the process characteristics of different systems. The modular neural network weight selection flow chart is shown in Figure 4 As shown; the network operation process is shown in Figure 5 .

[0075] Taking into account that the placement angles of components in different optical systems have different effects on the performance indicators of the optical systems, and at the same time, the types of commonly used optical systems are limited, therefore, in this application, different optical systems are classified, modular neural networks are trained for different types of optical systems, and the network weights are dynamically selected or adjusted according to different optical systems to ensure the prediction accuracy of the neural network.

[0076] In this application, deep neural networks (DNNs) can be used as a module in a modular neural network; it should be noted that it can also be replaced with other types of neural networks to achieve the same training effect.

[0077] S6. Obtaining an image quality degradation tolerance for the optical element based on the requirements of the process decision task, inputting the maximum image quality degradation tolerance as image quality degradation data into an optical element polishing process parameter prediction network model to obtain polishing process parameters that meet the image quality degradation tolerance;

[0078] The polishing process is not limited to a deterministic process, such as magnetorheological polishing, small grinding head polishing, ion beam polishing and other deterministic polishing processes;

[0079] The process parameters are not limited to one, such as machining gauge h, removal function RF, machining angle theta, machining dwell time dt, etc., that is, P = (RF, h, theta, dt);

[0080] The geometric optics evaluation index (image quality evaluation index) is not limited to one, such as the root mean square value of the point diagram, the modulation value at the cutoff frequency of the amplitude modulation function, etc. For example, for one of the optical elements of a system, Figure 2 The three process parameters of the removal function are simulated and calculated to obtain the corresponding virtual components and input them into the optical design software to simulate the optical system image quality corresponding to each virtual component. The image quality clarity evaluation results of a system are shown in Figure 6 .

[0081] The advantage of the present invention is that it solves the technical problem of traditional process parameters relying on experience-based adjustment through systematic data-driven modeling, and adopts a dynamic weight switching mechanism triggered by the optical system category to build a modular neural network. Although the method of the present invention requires a large amount of data collection in the early stage, once the neural network calculation model is built, the acquisition of decision-making process parameters is very fast and objective. The method of determining the deterministic polishing process parameters based on the image quality index of the neural network proposed in the present invention has obvious advantages in terms of calculation time and ease of understanding. Compared with the traditional debugging method, the process development efficiency can be increased by 5-6 times, which is particularly suitable for the precision processing scenarios of complex optical components such as aspheric surfaces.

[0082] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved. This is not limited herein.

[0083] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A method for determining polishing process parameters based on image quality indicators based on a neural network, characterized by: The specific steps include: S1. Determine the structural parameters of the optical system according to specific needs; the structural parameters of the optical system include the type of optical system, as well as the relative positions, placement angles, component surface shapes, and substrate materials of all optical components within the optical system; S2. Marking the optical element that needs to make a process decision in the optical system and recording the historical processing data of the optical element; the historical processing data includes process parameters, environmental parameters, and manufacturing error data corresponding to the process parameters; S3. Build the optical system, place the marked optical element in step S2 in the optical system, debug to obtain the actual image quality index, and collect data on the degree of deterioration compared with the ideal system image quality; S4 repeats steps S2 ~ S3 to obtain a data set of process parameters of the optical element and a data set of image quality deterioration; S5. Using the image quality deterioration dataset as the input dataset and the component process parameters as the output dataset, the dataset is input into a modular convolutional neural network for training to obtain a network model for predicting optical component polishing process parameters; The modular convolutional neural network includes a backbone convolution layer and a dynamic weight layer; the backbone convolution layer is used to extract features of image quality deterioration data; the dynamic weight layer dynamically loads different weights according to the optical system type, and is used to map features to the process parameter space; The training optimization process is as follows: S501. Perform weighted search based on the optical system feature vector. The weight index k is determined by the following formula: Where x is the optical system eigenvector; The function is used to find the input parameter value that makes the function reach the maximum value, or return the index of the maximum value element in the data set; S502. Input the image quality deterioration data into the modular convolutional neural network; the backbone convolution layer extracts the feature F of the image quality deterioration data: Among them, F is the image quality deterioration data feature extracted by the backbone convolution layer, and is also the input of the dynamic weight layer; Conv represents the convolution function; The feature F of the image quality deterioration data output by the backbone convolution layer is flattened into a vector f, and the dynamically loaded weights PT k Map to the process parameter space and calculate the predicted value of the process parameter; the calculation formula is as follows: in, is the ReLU activation function, Bias term, is the predicted value of process parameters; S503. Using the difference between the predicted process parameters and the actual process parameters as a loss function; S6. According to the requirements of the process decision task, the image quality deterioration tolerance of the optical element is obtained, and the maximum value of the image quality deterioration tolerance is input as the image quality deterioration data into the optical element polishing process parameter prediction network model to obtain the polishing process parameters that meet the image quality deterioration tolerance.

2. The method for determining polishing process parameters based on image quality indicators based on a neural network according to claim 1, characterized in that: The process parameter data in step S2 include machining track gauge h, removal function RF, machining angle theta, and machining dwell time dt.

3. The method for determining polishing process parameters based on image quality indicators based on a neural network according to claim 1, characterized in that: In step S4, steps S2 to S3 are repeated for at least N groups, where N is greater than or equal to 50.

4. The method for determining polishing process parameters based on image quality indicators based on a neural network according to claim 1, characterized in that: The loss function of step S503 is expressed as follows: Where, h i Indicates the actual processing gauge, represents the predicted processing trajectory, Indicates the actual processing angle, represents the predicted processing angle, … represents other process parameters.

5. The method for determining polishing process parameters based on image quality indicators based on a neural network according to claim 4, characterized in that: In the training optimization process of step S5, the Adam algorithm is used as the optimization algorithm to fit the relationship between the input data set and the output data set under different optical systems, and the weight files trained by different optical systems are numbered and recorded as PT k , ; where k is the number of the optical system and W is the number of types of optical systems.

Images