Spray disc cleaning process optimization method based on adaptive learning and multi-parameter modeling

Abstract

The application provides a spray disc cleaning process optimization method based on adaptive learning and multi-parameter modeling, and belongs to the technical field of semiconductor processing based on deep learning; process parameters and spray disc surface images in a spray disc surface cleaning process are collected; a cleaning quality prediction model is constructed, the collected data are taken as input, and cleaning cleanliness and residual particle degree prediction results are output; a spray disc surface cleaning process multi-objective optimization problem is established, a multi-objective evolutionary algorithm is used in combination with the constructed cleaning quality prediction model to obtain a Pareto optimal spray disc surface cleaning process parameter solution set; a spray disc surface cleaning process multi-parameter optimization method based on reinforcement learning is designed, an optimal spray disc surface cleaning process parameter vector is obtained, and the vector is applied to an actual spray disc surface cleaning process flow. The application can realize dynamic adjustment and intelligent optimization of cleaning parameters without excessive dependence on artificial experience, and is suitable for different product working conditions and pollution characteristics.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of semiconductor processing based on deep learning, and particularly relates to a spraying disc cleaning process optimization method based on adaptive learning and multi-parameter modeling. BACKGROUND

[0002] The inner electrode is also called a shower head, a spraying disc, or a plasma shunt disc, and is one of the most core components and consumables in the cavity of a dry etching machine. When in operation, radio frequency electric fields are applied to both ends of the electrode, process gas is uniformly dispersed through the micropores and is ionized at the same time, and finally reaches the wafer surface to complete etching. In advanced process nodes, surface cleanliness has a significant impact on device performance, and the stability and efficiency of the cleaning process have become important indicators of manufacturing capacity. However, the cleaning of the spraying disc surface involves multiple physical and chemical parameters such as temperature, time, chemical ratio, ultrasonic frequency, and spraying pressure, and the coupling between variables is complex, resulting in high difficulty in process control. In recent years, the development of artificial intelligence and machine learning technology has provided a new path for intelligent modeling and parameter optimization of the cleaning process. Based on historical process data and cleanliness feedback, a prediction model can be constructed to accurately associate parameters and cleaning results. Further combined with adaptive optimization algorithms, dynamic parameter adjustment and recommendation can also be achieved, breaking through the limitations of traditional settings and improving the intelligence, flexibility, and robustness of the cleaning process.

[0003] The existing semiconductor spraying disc surface cleaning method mainly includes the following three types:

[0004] RCA (Radio Corporation of America Cleaning) cleaning method: This method was first proposed by the RCA company in the United States. This method mainly consists of two steps: S1, a mixture of ammonia, hydrogen peroxide, and deionized water is used to remove organic contamination and particulate matter on the surface of the spraying disc; S2, a mixture of hydrochloric acid, hydrogen peroxide, and deionized water is used to remove metal ion contamination. In actual processes, operators usually perform batch cleaning according to fixed ratios and temperature and time conditions. However, this method has obvious limitations: the cleaning solution is highly corrosive and may damage the spraying disc surface material; the process flow is complex, and the control of cleaning parameters relies on human experience, lacking dynamic optimization capability; when facing different types of contamination or new materials, the cleaning effect is unstable. Therefore, RCA cleaning is difficult to meet the dual requirements of stability and adaptability in response to the cleaning needs of high-density integrated circuits and advanced process nodes.

[0005] Ultrasonic / megasonic cleaning method: this method is a kind of physical cleaning method which uses acoustic cavitation effect to remove particle contamination. In the cleaning process, the surface of the spray tray is immersed in the cleaning solution, and the sound wave propagates in the liquid and produces micro bubbles. When these bubbles burst on the surface of the spray tray, they release energy, which can remove micrometer and submicron particles. However, this type of cleaning method is very sensitive to parameters such as sound wave frequency, power density, and action time. Different materials and structures of the spray tray surface have different tolerances to cleaning intensity. Too strong may cause the film layer to peel off, and too weak may not achieve good cleaning effect. At present, industrial parameters are often set by relying on experience, and there is a lack of unified evaluation and optimization mechanism, making it difficult to achieve targeted regulation, resulting in poor cleaning consistency and narrow process window.

[0006] Dry cleaning method: this method is a kind of cleaning technology that relies on plasma, ozone or ultraviolet light to perform gas phase treatment on the surface of the spray tray, which is suitable for removing organic contaminants and some metal residues. Common operations include introducing oxygen into the plasma environment to generate active oxygen species that react with organic matter on the surface of the spray tray to form volatile byproducts. However, dry cleaning has limited cleaning capacity and slow processing rate, making it difficult to effectively remove larger particles or inorganic contaminants. At the same time, high-energy plasma can easily damage the microstructure, and improper parameter control can cause material modification or etching. In addition, this method requires high precision equipment and gas control, and the cost of process development is high. Moreover, it cannot adjust parameters in real time according to the type of pollution. Therefore, dry cleaning has certain contradictions between ensuring universality and high cleanliness, and lacks self-adaptive optimization ability for different working conditions and product structures.

[0007] Therefore, the existing methods often set parameters with fixed values or limited ranges, lack dynamic adaptability to different product types and pollution characteristics, rely too much on expert experience, have low automation level, and consume high labor cost and chemical reagent cost. SUMMARY

[0008] To solve the above problems, the present application proposes a spray tray surface cleaning process optimization method that combines adaptive learning and multi-parameter collaborative modeling. This method can dynamically adjust and intelligently optimize cleaning parameters without relying too much on human experience, adapt to different product conditions and pollution characteristics, improve overall process yield and cleanliness compliance rate, and promote the evolution of the semiconductor manufacturing process to a higher level of intelligence.

[0009] The present application specifically provides a spray tray surface cleaning process optimization method based on adaptive learning and multi-parameter modeling, including the following processes:

[0010] S1, collecting process parameter data and spray tray surface image data during the spray tray surface cleaning process; the process parameter data includes spray tray surface cleaning liquid temperature data, spray tray surface cleaning time data, spray tray surface cleaning liquid ratio, spray tray surface cleaning liquid flow rate, and the like;

[0011] S2, constructing a spray tray surface cleaning quality prediction model, taking the data collected in S1 as input, and outputting spray tray surface cleaning cleanliness and spray tray surface residual particle degree prediction results; the spray tray surface cleaning quality prediction model includes a spray tray surface cleaning process data feature extraction module, a spray tray surface image feature extraction module, and a spray tray surface cleaning quality prediction module;

[0012] S3, taking maximizing spray tray surface cleaning cleanliness and spray tray surface residual particle degree as an objective function, taking spray tray surface cleaning process parameters as decision variables, establishing a spray tray surface cleaning process multi-objective optimization problem, and using a multi-objective evolutionary algorithm to solve the problem to obtain a Pareto optimal spray tray surface cleaning process parameter solution set in combination with the spray tray surface cleaning quality prediction model constructed in S2;

[0013] S4, designing a spray tray surface cleaning process multi-parameter optimization method based on reinforcement learning, taking the best solution in the Pareto optimal spray tray surface cleaning process parameter solution set in S3 as an environment, obtaining an optimal spray tray surface cleaning process parameter vector, and applying it to an actual spray tray surface cleaning process flow.

[0014] Preferably, the spray tray surface cleaning liquid temperature data is collected by a platinum resistance thermometer, and the collection time period is 1 second; The spray tray surface cleaning liquid ratio is collected by measuring the conductivity of the cleaning liquid using a conductivity sensor, and the cleaning liquid ratio is calculated in combination with a pre-set conductivity-ratio mapping relationship. The conductivity sensor collects data at a time period of 1 second. .

[0015] Preferably, the spray tray surface cleaning process data feature extraction module is used to extract multi-time scale features of process parameter data, and the module is composed of three time series neural networks of different time scales and a six-head attention mechanism. The module models multi-scale time features of time series of process parameter data during the cleaning process.

[0016] The spray tray surface image feature extraction module is used to extract spray tray surface enhanced spatial features. The module is composed of a multi-scale graph convolutional network, a channel attention mechanism layer, and a spatial self-attention mechanism layer. The module models spatial distribution features of particles, stains, and textures in the image of the cleaned spray tray surface.​​

[0017] The spray disc surface cleaning quality prediction module outputs a prediction result based on the input of the spray disc surface strengthening space features and the process parameter data multi-time scale features.

[0018] Preferably, the spray disc surface cleaning process data feature extraction module specifically comprises:

[0019] The spray disc surface cleaning process data sequence is input into a small-scale time sequence neural network to obtain process parameter data short-term time dependence features; the small-scale time sequence neural network comprises a convolution kernel scale of 1 convolution layer and a time step of 1 long short-term memory network layer, the convolution kernel scale of 1 convolution layer is used to extract local features of the spray disc surface cleaning process data sequence, and the time step of 1 long short-term memory network layer is used to model the local features in time sequence to obtain the process parameter data short-term time dependence features;

[0020] The spray disc surface cleaning process data sequence is input into a medium-scale time sequence neural network to obtain process parameter data short-term time dependence features; the medium-scale time sequence neural network comprises a convolution kernel scale of 3 convolution layer and a time step of 3 long short-term memory network layer, the convolution kernel scale of 3 convolution layer is used to extract local features of the spray disc surface cleaning process data sequence, and the time step of 3 long short-term memory network layer is used to model the local features in time sequence to obtain the process parameter data medium-term time dependence features;

[0021] The spray disc surface cleaning process data sequence is input into a large-scale time sequence neural network to obtain process parameter data short-term time dependence features; the large-scale time sequence neural network comprises a convolution kernel scale of 7 convolution layer and a time step of 7 long short-term memory network layer, the convolution kernel scale of 7 convolution layer is used to extract local features of the spray disc surface cleaning process data sequence, and the time step of 7 long short-term memory network layer is used to model the local features in time sequence to obtain the process parameter data long-term time dependence features;

[0022] The process parameter data short-term time dependence features, the process parameter data medium-term time dependence features and the process parameter data long-term time dependence features are input into a six-head attention mechanism layer for sequence-level feature fusion to obtain process parameter data multi-time scale features; the six-head attention mechanism layer comprises six self-attention mechanisms, which are respectively used to capture bidirectional interaction features of the process parameter data short-term time dependence features and the process parameter data medium-term time dependence features, the process parameter data short-term time dependence features and the process parameter data long-term time dependence features, and the process parameter data medium-term time dependence features and the process parameter data long-term time dependence features; the obtained bidirectional interaction features are summed in the sequence dimension to obtain the process parameter data multi-time scale features.

[0023] Preferably, the spray tray surface image feature extraction module is specifically:

[0024] The spray tray surface image data is input into the multi-scale graph convolution network to obtain enhanced space features for characterizing the spray tray surface particles, stains and textures; the multi-scale graph convolution network is composed of a first graph convolution layer, a second graph convolution layer, a third graph convolution layer and a fourth graph convolution layer, a first max-pooling layer, a second max-pooling layer, a batch normalization layer and a ReLU nonlinear activation function layer;

[0025] First, the spray tray surface image data is input into the first graph convolution layer, then input into the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation to obtain spray tray surface space features; second, the spray tray surface space features are input into the second graph convolution layer, then input into the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation to obtain spray tray surface shallow space features; third, the spray tray surface shallow space features are input into the first max-pooling layer for down-sampling to extract spray tray surface coarse-fine scale space structure to obtain down-sampling features; fourth, the first down-sampling features are input into the third graph convolution layer to obtain the spray tray surface middle layer space features after input into the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation; fifth, the spray tray surface middle layer space features are input into the fourth graph convolution layer to obtain the spray tray surface high layer space features after input into the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation; finally, the spray tray surface deep layer space features are input into the second max-pooling layer for down-sampling to extract the spray tray surface local-global scale space structure to obtain the spray tray surface deep layer space features;

[0026] The down-sampling features, the spray tray surface middle layer space features and the spray tray surface deep layer space features are respectively input into the channel attention mechanism to enhance the expression ability of the space features; first, the channel attention mechanism obtains the description vectors of the three features by global average pooling, then inputs the three vectors into the Sigmoid activation function to obtain the weight coefficients of the three vectors; second, the three features are respectively multiplied with the corresponding weight coefficients in the feature dimension to obtain the channel weighted down-sampling features, the channel weighted spray tray surface middle layer space features and the channel weighted spray tray surface deep layer space features;

[0027] The channel weighted down-sampling features, the channel weighted spray tray surface middle layer space features and the channel weighted spray tray surface deep layer space features are simultaneously input into the spatial self-attention mechanism to fuse the three features in the channel dimension to obtain the spray tray surface enhanced space features.

[0028] Preferably, the spray disc surface cleaning quality prediction module projects the spray disc surface strengthening spatial features into vectors and inputs them into a two-layer fully connected network with process parameter data multi-time scale features, both using ReLU activation and applying Dropout; outputs the spray disc surface cleaning cleanliness and the spray disc surface residual particle degree, both of which are numerical values between 0 and 1, 0 indicating poor spray disc surface cleaning cleanliness and a large number of spray disc surface residual particles, and 1 indicating high spray disc surface cleaning cleanliness and a small number of spray disc surface residual particles.

[0029] Preferably, the S3 specific process is:

[0030] S31, establishing a spray disc surface cleaning process multi-objective optimization problem, for the objective function, consisting of and , wherein is a spray disc surface cleaning process parameter vector, is a function of the spray disc surface cleaning cleanliness with respect to , and is the spray disc surface residual particle degree;

[0031] S32, using Latin hypercube sampling to sample in the spray disc surface cleaning process parameter vector space to obtain N groups of samples;

[0032] S33, using the cleaning quality prediction model of S2 as a surrogate model to predict the spray disc surface cleaning cleanliness and the spray disc surface residual particle degree of each solution in the current population, and then calculating the objective function value of each solution;

[0033] S34, using the objective function value of the current population to perform non-dominated sorting and crowding degree calculation to obtain the non-dominated sorting of the solutions;

[0034] S35, performing binary tournament selection on the current population to obtain a parent population;

[0035] S36, performing binary crossover and polynomial mutation operations on the parent population to obtain an offspring population;

[0036] S37, using the cleaning quality prediction model of S2 as a surrogate model to predict the spray disc surface cleaning cleanliness and the spray disc surface residual particle degree of each solution in the offspring population, and then calculating the objective function value of each solution;

[0037] S38, merging the parent population and the offspring population, using the objective function value of the merged population to perform non-dominated sorting and crowding degree calculation to obtain the non-dominated sorting of the solutions, and selecting solutions as the next generation population according to the non-dominated level from low to high;

[0038] S39, if the preset function evaluation times are reached, output the solution set of the spray disc surface cleaning process multi-objective optimization problem, otherwise return to step S33.

[0039] Preferably, the specific process of S4 is:

[0040] Define the basic elements of the spray disc surface cleaning process multi-parameter optimization method based on reinforcement learning, including state, action, reward, return and strategy;

[0041] Define the environment to provide the initial solution for reinforcement learning; the environment is determined by the solution set of the spray disc surface cleaning process multi-objective optimization problem obtained by S3, specifically, calculate the L2 norm between all solution target values in the solution set of the spray disc surface cleaning process multi-objective optimization problem and , and take the solution with the minimum L2 norm in the solution set as the environment;

[0042] Build a strategy network to learn the mapping relationship between state and action selection probability; the strategy network is composed of three layers of feedforward fully connected layers; input the current state into the strategy network, output the probability distribution of various actions, and select the action with the maximum probability as the action at the next time step;

[0043] Execute the reinforcement learning strategy to output the converged strategy network;

[0044] Input the ideal spray disc surface cleaning cleanliness and spray disc surface cleaning residual particle degree expected by the engineer at the current time step into the obtained converged strategy network to output the action probability distribution, and select the action with the maximum probability as the optimal action to adjust the current spray disc surface cleaning process parameters.

[0045] Preferably, the state is defined as the spray disc surface cleaning cleanliness and spray disc surface cleaning residual particle degree obtained by inputting the spray disc surface cleaning process parameter vector generated by the action into the spray disc surface cleaning quality prediction model; the action is defined as adjusting all parameter values in the spray disc surface cleaning process parameter vector within the spray disc surface cleaning process parameter design space; the reward is defined as the L2 norm between the state corresponding to the spray disc surface cleaning process parameter vector adjusted by the action and , the larger the L2 norm, the greater the reward; the return is defined as the average value of the cumulative reward in the future time steps; the strategy is defined as the probability of selecting a specific action under the current state, and the goal is to learn the optimal strategy to maximize the return;

[0046] Preferably, for each parameter in the spray disc surface cleaning process parameter vector, there are three actions: increasing the value by 1%, decreasing the value by 1%, and not adjusting the value.

[0047] Compared with the prior art, the innovation points of the present application include:

[0048] (1) The spray disc surface cleaning process is deeply integrated with image perception and time series modeling technology. The application innovatively constructs a spray disc surface cleaning quality prediction model, which combines a spray disc surface image feature extraction module to process spatial distribution features and a process data feature extraction module to process multi-time scale features. The image feature extraction module uses a multi-scale graph convolution network, a channel attention mechanism, and a spatial self-attention mechanism to perform intensive spatial modeling on the cleaned spray disc surface image; at the same time, the process data feature extraction module uses three different scale time series neural networks and a six-head attention mechanism to capture short-term, medium-term, and long-term dependencies of parameters such as temperature, time, and ratio. This multi-modal fusion avoids the shortcomings of empirical rules and achieves accurate prediction of cleaning effect.

[0049] (2) A multi-parameter optimization framework based on reinforcement learning. The state is defined as the cleaning cleanliness and the residual particle degree, the action is defined as the up-regulation, down-regulation or no adjustment of each process parameter, and the reward is based on the L2 norm maximization of the state and the ideal target. It is worth noting that this method uses the Pareto optimal solution set as the initial environment to accelerate the convergence of the strategy network.

[0050] The beneficial effects brought by the innovation points of the application include:

[0051] (1) Improve the precision and automation level of the spray disc surface cleaning process: by combining image perception and time series modeling, the model's perception ability for complex contamination patterns and process changes is enhanced, avoiding the shortcomings of traditional methods that rely on empirical rules, and significantly improving the accuracy of cleaning effect prediction;

[0052] (2) Enhance the adaptability and flexibility of the cleaning process: by combining the reinforcement learning algorithm, the system can dynamically adjust the cleaning parameters in real time according to feedback, adapt to different contamination levels and spray disc surface materials, reduce manual intervention, and improve the automation level of the production line;

[0053] (3) Realize multi-objective optimization of cleaning process design: by combining the group optimization algorithm and reinforcement learning, the group optimization algorithm provides high-quality spray disc surface cleaning process parameter solutions for reinforcement learning, making the reinforcement learning converge faster and obtaining the global optimal spray disc surface cleaning process parameter solution in a relatively short time. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The overall technical route flowchart of the application.

[0055] Figure 2 The spray disc surface cleaning quality prediction model structure diagram of the application.

[0056] Figure 3 The reinforcement learning flowchart of the application.

[0057] Figure 4 This is a graph showing the accuracy versus loss function of the spray plate surface cleaning quality prediction model of the present invention.

[0058] Figure 5 This is a schematic diagram of the solution set of Pareto optimal spray disc surface cleaning process parameters of the present invention.

[0059] Figure 6 This is an analysis diagram of the surface cleaning process of the spray disc based on reinforcement learning optimization according to the present invention. Detailed Implementation

[0060] This invention proposes an optimization method for the surface cleaning process of spray discs based on adaptive learning and multi-parameter modeling, such as... Figure 1 As shown, the overall process includes: S1, collecting process parameter data and spray tray surface image data during the spray tray surface cleaning process; the process parameter data includes spray tray surface cleaning liquid temperature data, spray tray surface cleaning time data, spray tray surface cleaning liquid ratio, spray tray surface cleaning liquid, and spray tray surface cleaning liquid flow rate; S2, inputting the data collected in S1 into the trained spray tray surface cleaning quality prediction model, and outputting the prediction results of the spray tray surface cleaning cleanliness and the degree of residual particles on the spray tray surface; the spray tray surface cleaning quality prediction model includes a spray tray surface cleaning process data feature extraction module, a spray tray surface image feature extraction module, and a spray tray surface cleaning quality prediction module; the spray tray surface cleaning process data feature extraction module is used to extract the temporal features of the spray tray surface cleaning process data, including three temporal neural networks at different time scales and a six-head attention mechanism, aiming to perform multi-scale temporal feature modeling of the process parameter data time series during the cleaning process; the spray... The spray tray surface image feature extraction module extracts the cleaning state features of the spray tray surface, including a multi-scale graph convolutional network, a channel attention mechanism layer, and a spatial self-attention mechanism layer, aiming to model the spatial distribution features of particles, stains, and textures in the cleaned spray tray surface image. In step S3, using the maximization of the cleanliness of the spray tray surface and the degree of residual particles on the spray tray surface as the objective function, and taking the spray tray surface cleaning process parameters described in step S1 as decision variables, a multi-objective optimization problem for the spray tray surface cleaning process is established. The Pareto optimal solution set of the spray tray surface cleaning process parameters is obtained by combining a multi-objective evolutionary algorithm with the spray tray surface cleaning quality prediction model established in step S2. In step S4, a reinforcement learning-based multi-parameter optimization method for the spray tray surface cleaning process is designed to obtain the optimal spray tray surface cleaning process parameter vector and applied to the actual spray tray surface cleaning process. This technical framework effectively improves the stability, accuracy, and cleanliness compliance rate of the semiconductor spray tray surface cleaning process through multi-level intelligent decision-making.

[0061] The invention will be further described below with reference to specific embodiments.

[0062] I. Constructing a semiconductor shower plate surface cleaning multi-modal dataset

[0063] First, the process parameter data and shower plate surface image data during the shower plate surface cleaning process are systematically collected as follows:

[0064] Shower plate surface cleaning process parameter data collection: Collecting process parameter data of the whole process of shower plate surface cleaning includes shower plate surface cleaning liquid temperature data , shower plate surface cleaning time data , shower plate surface cleaning liquid ratio , shower plate surface cleaning liquid , and shower plate surface cleaning liquid flow . These data are sampled at high frequency through deployment, ensuring the time sequence integrity and accuracy of the data, facilitating the reflection of the real process change process; shower plate surface cleaning liquid temperature data is collected by platinum resistance thermometer, with a collection time period of ; shower plate surface cleaning liquid ratio is collected by conductivity sensor, with a collection time period of ; shower plate surface cleaning liquid is collected by pH detector, with a collection time period of ; shower plate surface cleaning liquid flow is collected by Coriolis flowmeter, with a collection time period of .

[0065] Shower plate surface image data collection: For each batch of cleaned shower plate surface sample, high-resolution microscope is used to shoot shower plate surface image, to obtain image data of shower plate surface microscopic particle distribution and pollution residue. During the collection process, laser positioning system is used to ensure the positioning uniformity of the sample, and the image resolution is set to pixels, to ensure the consistency of data between different samples, thereby supporting accurate comparison and analysis of image features.

[0066] Shower plate surface cleaning multi-modal data preprocessing: First, due to the different sampling frequencies of different sensors, the obtained shower plate surface cleaning liquid temperature data , shower plate surface cleaning liquid ratio , shower plate surface cleaning liquid , and shower plate surface cleaning liquid flow In order to carry out alignment in the time dimension, the Kalman filter is used to interpolate and complete four kinds of modal data to align the time dimension; secondly, since there is uneven illumination environmental noise in the spray disc surface cleaning process, the collected spray disc surface image data needs to be denoised, and the Gaussian filter is used to denoise the spray disc surface image data to eliminate the uneven illumination noise.

[0067] II. Constructing a spray disc surface cleaning quality prediction model

[0068] The spray disc surface cleaning quality prediction model structure is shown in Figure 2 The model includes a spray disc surface cleaning process data feature extraction module, a spray disc surface image feature extraction module and a spray disc surface cleaning quality prediction module; the spray disc surface cleaning process data feature extraction module and the spray disc surface image feature extraction module respectively extract features from the collected spray disc surface cleaning process parameter data, spray disc surface image data and spray disc surface cleaning effect data.

[0069] 1. Spray disc surface cleaning process data feature extraction module

[0070] This module is composed of three different time scale time series neural networks and six attention mechanisms, aiming to model multi-scale time features of the time series of process parameter data in the cleaning process. The spray disc surface cleaning process data sequence is input into the cleaning process data feature extraction module, and the specific steps are as follows:

[0071] The spray disc surface cleaning process data sequence is input into the small-scale time series neural network to obtain the short-term time dependence feature of the process parameter data; the small-scale time series neural network includes a convolution kernel scale of 1 convolution layer and a time step of 1 long short-term memory network layer, the convolution kernel scale of 1 convolution layer is used to extract the local features of the spray disc surface cleaning process data sequence, and the time step of 1 long short-term memory network layer is used to model the local features to obtain the short-term time dependence feature of the process parameter data;

[0072] The spray disc surface cleaning process data sequence is input into the medium-scale time series neural network to obtain the short-term time dependence feature of the process parameter data; the medium-scale time series neural network includes a convolution kernel scale of 3 convolution layer and a time step of 3 long short-term memory network layer, the convolution kernel scale of 3 convolution layer is used to extract the local features of the spray disc surface cleaning process data sequence, and the time step of 3 long short-term memory network layer is used to model the local features to obtain the medium-term time dependence feature of the process parameter data;

[0073] The process data sequence of the spray tray surface cleaning process is input into a large-scale temporal neural network to obtain the short-term time dependence features of the process parameter data. The large-scale temporal neural network includes a convolutional layer with a kernel size of 7 and a long short-term memory network layer with a time step of 7. The convolutional layer with a kernel size of 7 is used to extract local features of the spray tray surface cleaning process data sequence, and the long short-term memory network layer with a time step of 7 is used to perform temporal modeling on the local features to obtain the long-term time dependence features of the process parameter data.

[0074] The short-term, medium-term, and long-term time dependence features of process parameter data are input into a six-head attention mechanism layer for sequence-level feature fusion to obtain multi-timescale features of the process parameter data. The six-head attention mechanism layer includes six self-attention mechanisms, which are used to capture the bidirectional interaction features between the short-term and medium-term time dependence features of the process parameter data, the bidirectional interaction features between the short-term and long-term time dependence features of the process parameter data, and the bidirectional interaction features between the medium-term and long-term time dependence features of the process parameter data. The obtained bidirectional interaction features are summed along the sequence dimension to obtain the multi-timescale features of the process parameter data.

[0075] 2. Spray disc surface image feature extraction module

[0076] This module consists of a multi-scale graph convolutional network, a channel attention mechanism layer, and a spatial self-attention mechanism layer. It aims to model the spatial distribution characteristics of particles, stains, and textures in the surface image of the cleaned spray tray. The spray tray surface image data is input into the spray tray surface image data feature extraction module, and the specific steps are as follows:

[0077] 1) input the spray tray surface image data into the multi-scale graph convolution network to obtain enhanced spatial features for representing the spray tray surface particles, stains and textures; the multi-scale graph convolution network is composed of a first graph convolution layer, a second graph convolution layer, a third graph convolution layer and a fourth graph convolution layer, a first max pooling layer, a second max pooling layer, a batch normalization layer and a ReLU nonlinear activation function layer; the channel numbers of the four graph convolution layers are 32, 64, 128 and 256 respectively, and the node aggregation scales are 1, 2, 3 and 4 respectively; first, input the spray tray surface image data into the first graph convolution layer, then input the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation to obtain the spray tray surface spatial features; second, input the spray tray surface spatial features into the second graph convolution layer, then input the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation to obtain the spray tray surface shallow layer spatial features; third, input the spray tray surface shallow layer spatial features into the first max pooling layer for down-sampling to extract the spray tray surface coarse-fine scale spatial structure to obtain the down-sampled features; fourth, input the first down-sampled features into the third graph convolution layer to obtain the spray tray surface middle layer spatial features, then input the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation to obtain the spray tray surface middle layer spatial features; fifth, input the spray tray surface middle layer spatial features into the fourth graph convolution layer to obtain the spray tray surface high layer spatial features, then input the batch normalization layer and the ReLU nonlinear activation function layer for nonlinear transformation to obtain the spray tray surface high layer spatial features; finally, input the spray tray surface deep layer spatial features into the second max pooling layer for down-sampling to extract the spray tray surface local-global scale spatial structure to obtain the spray tray surface deep layer spatial features;

[0078] 2) input the down-sampled features, the spray tray surface middle layer spatial features and the spray tray surface deep layer spatial features into the channel attention mechanism respectively to enhance the expression ability of the spatial features and improve the sensitivity of the model to key pollution areas and multi-channel heterogeneous features; first, the channel attention mechanism obtains the description vectors of the three features through global average pooling, then inputs the three vectors into the Sigmoid activation function to obtain the weight coefficients of the three vectors; second, the three features are respectively multiplied with the corresponding weight coefficients in the feature dimension to obtain the channel weighted down-sampled features, the channel weighted spray tray surface middle layer spatial features and the channel weighted spray tray surface deep layer spatial features;

[0079] 3) input the channel weighted down-sampled features, the channel weighted spray tray surface middle layer spatial features and the channel weighted spray tray surface deep layer spatial features into the spatial self-attention mechanism to fuse the three features in the channel dimension to obtain the spray tray surface enhanced spatial features.

[0080] 3. Spray tray surface cleaning quality prediction module

[0081] The spray disc surface cleaning quality prediction module outputs the spray disc surface cleaning degree and the spray disc surface residual particle degree prediction results; the spray disc surface strengthening space features are flattened into vectors and input into a two-layer fully connected network with process parameter data multi-time scale features, the number of hidden units is 256 and 64 respectively, ReLU activation is adopted, and Dropout (p=0.3) is applied; the output is the spray disc surface cleaning degree and the spray disc surface residual particle degree, wherein the spray disc surface cleaning degree and the spray disc surface residual particle degree are both numerical values between 0 and 1, 0 represents poor spray disc surface cleaning degree and a large number of spray disc surface residual particles; 1 represents high spray disc surface cleaning degree and a small number of spray disc surface residual particles.

[0082] III. Designing a Pareto optimal process parameter search strategy to obtain a Pareto optimal solution set

[0083] Designing a Pareto optimal process parameter search strategy to obtain a Pareto optimal solution set provides an initial solution for spray disc surface process parameter adaptive optimization, as follows:

[0084] A spray disc surface cleaning process multi-objective optimization problem is constructed; the present application takes the maximum spray disc surface cleaning degree and the spray disc surface residual particle degree as the objective function, and the spray disc surface cleaning process parameters in S1 as the decision variables to establish a spray disc surface cleaning process multi-objective optimization problem, as follows:

[0085]

[0086] Wherein The spray disc surface cleaning process parameter vector, ; The spray disc surface cleaning degree is a function of , and The spray disc surface residual particle degree is a function of ; The objective function is composed of and Because the spray disc surface cleaning degree and the spray disc surface residual particle degree are maximum indicators, the inverse is taken in the objective function.

[0087] Solving the spray disc surface cleaning process multi-objective optimization problem, the specific steps are as follows:

[0088] 1) Latin hypercube sampling is adopted to sample in the spray disc surface cleaning process parameter vector space to obtain N groups of samples, and the N groups of samples are used as the initial population of a multi-objective evolutionary algorithm (MOEA);

[0089] 2) using the cleaning quality prediction model as a surrogate model to predict the spray disc surface cleaning cleanliness and the spray disc surface residual particle degree of each solution in the current population, and then calculating the objective function value of each solution;

[0090] 3) using the objective function value of the current population to perform non-dominated sorting and crowdedness calculation to obtain the non-dominated sorting of the solution;

[0091] 4) performing binary tournament selection on the current population to obtain the parent population;

[0092] 5) performing binary crossover and polynomial mutation operations on the parent population to obtain the offspring population;

[0093] 6) using the cleaning quality prediction model as a surrogate model to predict the spray disc surface cleaning cleanliness and the spray disc surface residual particle degree of each solution in the offspring population, and then calculating the objective function value of each solution;

[0094] 7) merging the parent population and the offspring population, using the objective function value of the merged population to perform non-dominated sorting and crowdedness calculation to obtain the non-dominated sorting of the solution, and selecting solutions as the next generation population from low to high according to the non-dominated level;

[0095] If the preset function evaluation number is reached, output the solution set of the spray disc surface cleaning process multi-objective optimization problem, otherwise return to step 2.

[0096] Four, constructing a spray disc surface cleaning process multi-parameter optimization method based on reinforcement learning

[0097] The spray disc surface cleaning process multi-parameter optimization method based on reinforcement learning realizes the adaptive optimization of spray disc surface cleaning process parameters, as shown in Figure 3 , specifically as follows:

[0098] Define the basic elements of the spray disc surface cleaning process multi-parameter optimization method based on reinforcement learning, including state , action , reward , return and policy ; the state is defined as the spray disc surface cleaning cleanliness and the spray disc surface residual particle degree obtained by inputting the spray disc surface cleaning process parameter vector generated by the action into the spray disc surface cleaning quality prediction model; the action is defined as adjusting all parameter values in the spray disc surface cleaning process parameter vector within the spray disc surface cleaning process parameter design space, and for each parameter in the spray disc surface cleaning process parameter vector, there are three actions of increasing the value by 1%, decreasing the value by 1% and not adjusting the value, a total of 243 actions; the reward L2 norm between the state corresponding to the spray disc surface cleaning process parameter vector adjusted by the action and , the larger the L2 norm, the greater the reward; the return is defined as the cumulative reward of the future time steps; the policy is defined as the probability of selecting a specific action in the current state, and the goal is to learn the optimal policy that maximizes the return.

[0099] The environment is defined as the initial solution for reinforcement learning; the environment is determined by the set of solutions obtained for the spray disc surface cleaning process multi-objective optimization problem, specifically the L2 norm between all solution objective values in the set of solutions for the spray disc surface cleaning process multi-objective optimization problem and , the solution with the smallest L2 norm in the solution set is taken as the environment.

[0100] The policy network is constructed to learn the mapping relationship between the state and the action selection probability; the policy network consists of three layers of feedforward fully connected layers; the current state is input into the policy network, and the probability distribution of 243 actions is output, and the action with the highest probability is selected as the action for the next time step.

[0101] The reinforcement learning policy is executed to output the optimal spray disc surface cleaning process parameter vector, specifically as follows:

[0102] 1) Use the decision variables of the defined environment initial solution as the starting state ;

[0103] 2) Input the current state into the policy network and output the action probability distribution;

[0104] 3) Select and execute the action and calculate the reward;

[0105] 4) Use the greedy algorithm to obtain the new state ;

[0106] 5) Update the policy network weights using the policy gradient, with the objective function being to maximize the return, specifically as follows:

[0107]

[0108] where is the weight of the policy network, is the maximum return, is the policy gradient method Actor-Critic;

[0109] 6) Repeat steps 2) to 5) until the preset time steps are met, and output the converged policy network.​

[0110] The convergence strategy network outputs the action probability distribution by inputting the ideal surface cleanliness of the spray tray and the degree of residual particles on the surface of the spray tray at the current time step. The action with the highest probability is selected as the optimal action to adjust the current surface cleaning process parameters of the spray tray. The adjusted surface cleaning process parameters are configured in the surface cleaning process of the spray tray to improve the surface cleaning quality of the spray tray.

[0111] V. Analysis of Experimental Results

[0112] The collected data on the surface cleaning process parameters, surface images, and cleaning effect of the spray tray were input into the proposed surface cleaning quality prediction model for the spray tray. Experimental simulations were then conducted, and the model's performance was evaluated using accuracy and loss curves. The experimental results are as follows: Figure 4 As shown:

[0113] Figure 4 The graphs show the accuracy (left) and loss function (right) curves of the spray tray surface cleaning quality prediction model during training. During training, by continuously optimizing the model parameters, the accuracy of the spray tray surface cleaning quality prediction was improved, while the model's loss value was reduced. The horizontal axis represents the number of training epochs, ranging from 1 to 50, showing the model's performance at different training stages. Figure 4 In (a), the accuracy gradually improves as training progresses. In the early stages of training, the accuracy increases slowly, but as the number of training rounds increases, the improvement in accuracy gradually accelerates and approaches 1, indicating that the model gradually converges and can make more accurate predictions about the surface cleaning quality of the spray tray. Figure 4 In (b), the loss function gradually decreases with increasing training epochs, indicating that the model reduces prediction errors and improves prediction ability through optimization. The loss decreases rapidly in the early stages of training, while the rate of decrease slows down in later stages, showing the gradual convergence of the model. At the 50th epoch, the model's prediction accuracy reaches 94.5%, and the loss function value is 0.375774, indicating that the model has high prediction accuracy and effectiveness, and can provide reliable support for predicting the surface cleaning quality of the spray tray.

[0114] The Pareto optimal front of the spray disc surface cleaning process parameters obtained by solving the multi-objective evolutionary algorithm combined with the established spray disc surface cleaning quality prediction model is as follows: Figure 5 As shown:

[0115] In the figure, the abscissa represents the fitness value of the cleaning cleanliness of the shower tray surface, and the ordinate represents the fitness value of the residual particle degree of the shower tray surface. The crowding distance distribution between the dominated points in the figure is relatively uniform, while maintaining a high diversity. This shows that the proposed optimization method can effectively maximize the cleaning cleanliness of the shower tray surface and minimize the residual particle degree of the shower tray surface, and successfully solve a set of Pareto optimal shower tray surface cleaning process parameter solution set.

[0116] The experimental verification of the shower tray surface cleaning process multi-parameter optimization method based on reinforcement learning shows that the effectiveness of the shower tray surface cleaning process self-adaptive optimization process. The experimental results are shown in Figure 6

[0117] Figure 6 From left to right, the stability, accuracy and cleanliness compliance rate of the shower tray surface cleaning process affected by reinforcement learning are shown. As can be seen from the figure, with the increase of the number of iterations of reinforcement learning, the stability, accuracy and cleanliness compliance rate of the shower tray surface cleaning model are steadily improved. This shows that in the process of continuous training and optimization, the reinforcement learning model gradually adapts to the complexity of the shower tray surface cleaning process, and effectively improves the performance of these key indicators. In addition, the experimental results further verify that the model framework proposed in this study has high adaptability, and can automatically adjust the strategy according to different shower tray surface cleaning process parameters and environmental conditions to achieve the best cleaning effect. Therefore, the optimization method based on reinforcement learning can effectively improve the accuracy and stability of the shower tray surface cleaning process, and provides solid technical support for the automatic cleaning process in semiconductor manufacturing.

[0118] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various changes and modifications to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0119] Although the specific embodiments of the present application have been described above, they are not intended to limit the scope of protection of the present application. Those skilled in the art should understand that various modifications or variations made on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.​

Claims

1. A method for optimizing the spray disc cleaning process based on adaptive learning and multi-parameter modeling, characterized in that, Includes the following processes: S1, Collect process parameter data and spray plate surface image data during the spray plate surface cleaning process; the process parameter data includes spray plate surface cleaning fluid temperature data, spray plate surface cleaning time data, spray plate surface cleaning fluid ratio, and spray plate surface cleaning fluid... Value and flow rate of cleaning fluid on the surface of the spray tray; S2, construct a spray tray surface cleaning quality prediction model, using the data collected in S1 as input, and output the prediction results of the cleanliness of the spray tray surface and the degree of residual particles on the spray tray surface; the spray tray surface cleaning quality prediction model includes a spray tray surface cleaning process data feature extraction module, a spray tray surface image feature extraction module, and a spray tray surface cleaning quality prediction module. The specific feature extraction module for the surface cleaning process data of the spray disc is as follows: The process data sequence of the spray plate surface cleaning process is input into a small-scale temporal neural network to obtain short-term time-dependent features of the process parameter data. The small-scale temporal neural network includes a convolutional layer with a kernel size of 1 and a long short-term memory network layer with a time step of 1. The convolutional layer with a kernel size of 1 is used to extract local features of the spray plate surface cleaning process data sequence, and the long short-term memory network layer with a time step of 1 is used to perform temporal modeling on the local features to obtain short-term time-dependent features of the process parameter data. The process data sequence of the spray plate surface cleaning process is input into a mesoscale temporal neural network to obtain the mid-term time dependence features of the process parameter data; The mesoscale temporal neural network includes a convolutional layer with a kernel size of 3 and a long short-term memory network layer with a time step of 3. The convolutional layer with a kernel size of 3 is used to extract local features of the spray plate surface cleaning process data sequence, and the long short-term memory network layer with a time step of 3 is used to perform temporal modeling on the local features to obtain the mid-term time dependence features of the process parameter data. The process data sequence of the spray tray surface cleaning process is input into a large-scale temporal neural network to obtain the long-term time dependence features of the process parameter data. The large-scale temporal neural network includes a convolutional layer with a kernel size of 7 and a long short-term memory network layer with a time step of 7. The convolutional layer with a kernel size of 7 is used to extract local features of the spray tray surface cleaning process data sequence, and the long short-term memory network layer with a time step of 7 is used to perform temporal modeling on the local features to obtain the long-term time dependence features of the process parameter data. The short-term, medium-term, and long-term time-dependent features of process parameter data are input into a six-head attention mechanism layer for sequence-level feature fusion to obtain multi-timescale features of process parameter data. The six-head attention mechanism layer includes six self-attention mechanisms, which are used to capture the bidirectional interaction features of short-term time dependence features and medium-term time dependence features of process parameter data, the bidirectional interaction features of short-term time dependence features and long-term time dependence features of process parameter data, and the bidirectional interaction features of medium-term time dependence features and long-term time dependence features of process parameter data, respectively. The obtained bidirectional interaction features are summed along the sequence dimension to obtain multi-timescale features of the process parameter data; S3 uses the objective function of maximizing the cleanliness of the spray tray surface and minimizing the residual particle level on the spray tray surface, and takes the spray tray surface cleaning process parameters as decision variables to establish a multi-objective optimization problem of the spray tray surface cleaning process. The Pareto optimal solution set of the spray tray surface cleaning process parameters is obtained by using a multi-objective evolutionary algorithm combined with the spray tray surface cleaning quality prediction model constructed in S2. S4. A multi-parameter optimization method for the spray plate surface cleaning process based on reinforcement learning is designed. The best solution in the Pareto optimal solution set of the spray plate surface cleaning process parameters in S3 is used as the environment to obtain the optimal spray plate surface cleaning process parameter vector and apply it to the actual spray plate surface cleaning process.

2. The method for optimizing the spray plate cleaning process based on adaptive learning and multi-parameter modeling as described in claim 1, characterized in that: Temperature data of the cleaning fluid on the spray tray surface was collected using a platinum resistance thermometer, with a collection period of [time period missing]. The method for collecting the cleaning fluid ratio on the spray plate surface is to measure the conductivity of the cleaning fluid using a conductivity sensor, and calculate the cleaning fluid ratio based on a preset conductivity-ratio mapping relationship. The conductivity sensor data collection period is specified. ; Spray plate surface cleaning solution Value by The detector collects data over a period of time. The flow rate of the cleaning fluid on the spray plate surface is collected by a Coriolis flow meter, with a collection period of [time period missing]. .

3. The method for optimizing the spray plate cleaning process based on adaptive learning and multi-parameter modeling as described in claim 1, characterized in that: The spray disc surface cleaning process data feature extraction module is used to extract multi-timescale features of the process parameter data. This module consists of three temporal neural networks with different time scales and a six-head attention mechanism to perform multi-scale time feature modeling of the process parameter data time series during the cleaning process. The spray plate surface image feature extraction module is used to extract and obtain the enhanced spatial features of the spray plate surface; This module consists of a multi-scale graph convolutional network, a channel attention mechanism layer, and a spatial self-attention mechanism layer, which models the spatial distribution features of particles, stains, and textures in the surface image of the cleaned spray tray; The spray plate surface cleaning quality prediction module outputs prediction results based on the input of the spatial characteristics of the spray plate surface strengthening and the multi-timescale characteristics of the process parameter data.

4. The method for optimizing the spray plate cleaning process based on adaptive learning and multi-parameter modeling as described in claim 3, characterized in that: The spray disc surface image feature extraction module specifically comprises: The surface image data of the spray tray is input into a multi-scale graph convolutional network to obtain enhanced spatial features for characterizing particles, stains and textures on the surface of the spray tray. The multi-scale graph convolutional network consists of a first graph convolutional layer, a second graph convolutional layer, a third graph convolutional layer and a fourth graph convolutional layer, a first max pooling layer, a second max pooling layer, a batching layer and a ReLU nonlinear activation function layer. First, the surface image data of the spray disk is input into a first convolutional layer, then into a batching layer and a ReLU nonlinear activation function layer for nonlinear transformation to obtain the spatial features of the spray disk surface. Second, the spatial features of the spray disk surface are input into a second convolutional layer, then into a batching layer and a ReLU nonlinear activation function layer for nonlinear transformation to obtain the shallow spatial features of the spray disk surface. Third, the shallow spatial features of the spray disk surface are input into a first max pooling layer for downsampling to extract the coarse-fine scale spatial structure of the spray disk surface to obtain downsampled features. Fourth, the downsampled features are input into a third convolutional layer, then into a batching layer and a ReLU nonlinear activation function layer for nonlinear transformation to obtain the mid-level spatial features of the spray disk surface. Fifth, the mid-level spatial features of the spray disk surface are input into a fourth convolutional layer, then into a batching layer and a ReLU nonlinear activation function layer for nonlinear transformation to obtain the high-level spatial features of the spray disk surface. Finally, the deep spatial features of the spray disk surface are input into a second max pooling layer for downsampling to extract the local-global scale spatial structure of the spray disk surface to obtain the deep spatial features of the spray disk surface. The downsampling features, mid-level spatial features of the spray disk surface, and deep spatial features of the spray disk surface are respectively input into the channel attention mechanism to enhance the expressive power of spatial features. First, the channel attention mechanism obtains the description vectors of the three features through global average pooling and then inputs them into the Sigmoid activation function to obtain the weight coefficients of the three vectors. Second, the three features are respectively dot-producted with their corresponding weight coefficients along the feature dimension to obtain the channel-weighted downsampling features, channel-weighted mid-level spatial features of the spray disk surface, and channel-weighted deep spatial features of the spray disk surface. The channel-weighted downsampling features, the channel-weighted mid-level spatial features of the spray disk surface, and the channel-weighted deep spatial features of the spray disk surface are simultaneously input into the spatial self-attention mechanism to perform channel-dimensional fusion of the three features to obtain the enhanced spatial features of the spray disk surface.

5. The method for optimizing the spray plate cleaning process based on adaptive learning and multi-parameter modeling as described in claim 3, characterized in that: The spray tray surface cleaning quality prediction module flattens the spatial features of the spray tray surface enhancement into a vector and inputs them along with the multi-timescale features of the process parameter data into a two-layer fully connected network. Both layers are activated using ReLU and Dropout is applied. The output is the surface cleanliness of the spray tray and the degree of residual particles on the surface of the spray tray. Both the surface cleanliness of the spray tray and the degree of residual particles on the surface of the spray tray are values ​​between 0 and 1. 0 indicates poor surface cleanliness and a large number of residual particles on the surface of the spray tray; 1 indicates high surface cleanliness and a small number of residual particles on the surface of the spray tray.

6. The method for optimizing the spray plate cleaning process based on adaptive learning and multi-parameter modeling as described in claim 1, characterized in that: The specific process of S3 is as follows: S31, Establish a multi-objective optimization problem for the surface cleaning process of the spray disc. The objective function is... and Composition, in which This is a vector of process parameters for cleaning the surface of the spray plate. The cleanliness of the spray tray surface is related to The function, The degree of residual particles on the surface of the spray tray; S32, N sets of samples are obtained by sampling the cleaning process parameter vector space on the surface of the spray plate using Latin hypercube sampling; S33, using the cleaning quality prediction model described in S2 as a surrogate model, predict the surface cleanliness of the spray tray and the degree of residual particles on the surface of the spray tray for each solution in the current population, and then calculate the objective function value for each solution; S34, using the objective function value of the current population to perform non-dominated sorting and crowding calculation to obtain the solution's non-dominated sorting; S35, perform a binary tournament selection on the current population to obtain the parent population; S36, perform binary crossover and polynomial mutation operations on the parent population to obtain the offspring population; S37, using the cleaning quality prediction model described in S2 as a surrogate model to predict the surface cleanliness of the spray plate and the degree of residual particles on the surface of the spray plate for each solution in the offspring population, and then calculating the objective function value for each solution; S38: Merge the parent and offspring populations. Utilize the objective function value of the merged population to perform non-dominated sorting and crowding calculations to obtain the solution's non-dominated sorting. Select from low to high non-dominated levels. Each solution serves as the next generation population; S39. If the preset number of function evaluations is reached, output the solution set of the multi-objective optimization problem of the spray plate surface cleaning process; otherwise, return to step S33.

7. The method for optimizing the spray plate cleaning process based on adaptive learning and multi-parameter modeling as described in claim 1, characterized in that: The specific process of S4 is as follows: Define the basic elements of a multi-parameter optimization method for spray disc surface cleaning process based on reinforcement learning, including state, action, reward, prize, and policy; Define the environment to provide an initial solution for reinforcement learning; The environment is determined by the solution set of the multi-objective optimization problem for the spray plate surface cleaning process obtained from S3. Specifically, it involves calculating the sum of the objective values ​​of all solutions in the solution set of the multi-objective optimization problem for the spray plate surface cleaning process. The L2 norm between the solutions is used to determine the environment of the solution set with the smallest L2 norm. Construct a mapping relationship between the policy network learning state and action selection probability; the policy network consists of three feedforward fully connected layers; input the current state into the policy network, output the probability distribution of various actions, and select the action with the highest probability as the action of the next time step; Execute the reinforcement learning policy and output the converged policy network; The convergence strategy network outputs the action probability distribution by inputting the ideal surface cleanliness of the spray tray and the degree of residual particles on the surface of the spray tray as inputs at the current time step. The action with the highest probability is selected as the optimal action to adjust the current surface cleaning process parameters of the spray tray.

8. The method for optimizing the spray disc cleaning process based on adaptive learning and multi-parameter modeling as described in claim 7, characterized in that: The state is defined as the surface cleanliness and residual particle level of the spray tray obtained by inputting the process parameter vector of the spray tray surface cleaning generated by the action into the spray tray surface cleaning quality prediction model. The action is defined as adjusting all parameter values ​​in the spray plate surface cleaning process parameter vector within the design space of the spray plate surface cleaning process parameters. The reward is defined as the state corresponding to the spray plate surface cleaning process parameter vector after the action adjustment. The L2 norm between the two values ​​is used; the larger the L2 norm, the greater the reward. The reward is defined as the future L2 norm. The cumulative average reward over time steps; the policy is defined as the probability of choosing a specific action in the current state, with the goal of learning the optimal policy to maximize the reward.

9. The method for optimizing the spray plate cleaning process based on adaptive learning and multi-parameter modeling as described in claim 8, characterized in that: For each parameter in the process parameter vector for cleaning the surface of the spray plate, there are three actions: increasing the value by 1%, decreasing the value by 1%, and not adjusting the value.

Images