Semiconductor valve sealing surface honing track optimization method and system based on spatial attention mechanism

Abstract

The application provides a semiconductor valve sealing surface honing track optimization method and system based on a spatial attention mechanism, and relates to the technical field of semiconductor equipment processing. The method comprises the following steps: constructing a sealing surface-process correlation database; training a preset convolutional neural network model according to the sealing surface-process correlation database to obtain a trained convolutional neural network model; inputting the topography data of a sealing surface to be processed into the trained convolutional neural network model, and positioning a key micro area on the sealing surface through the spatial attention mechanism thereof; and extracting a plurality of topography reference points corresponding to the key micro area from the sealing surface-process correlation database to form a topography reference point data set. The application realizes adaptive and accurate optimization of the honing track, and improves the processing quality of the sealing surface and the sealing performance of the valve.

Description

Technical Field

[0001] This invention relates to the field of semiconductor equipment processing technology, and in particular to a method and system for optimizing the honing trajectory of semiconductor valve sealing surfaces based on spatial attention mechanisms. Background Technology

[0002] Semiconductor valves are core components for high-purity fluid control in semiconductor processes. The quality of their sealing surface machining directly determines key performance characteristics such as internal leakage during helium detection and seal life. Furthermore, they are a significant factor affecting the yield of core semiconductor processes such as etching, deposition, and ion implantation. As the semiconductor industry moves towards higher processes and higher purity, the industry has placed stringent demands on the precision of valve sealing surface machining, adaptability to operating conditions, and long-term operational stability.

[0003] Currently, the honing process for semiconductor valve sealing surfaces in the industry generally faces the following technical defects: The honing trajectory optimization lacks spatial attention mechanisms for guidance; morphological characteristics of key micro-areas such as the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing band are not quantitatively analyzed; optimization is based solely on macroscopic sealing performance indicators; and process verification is limited to static testing under a single rated operating condition, without constructing a multi-condition dynamic simulation verification system. These defects lead to over- or under-honing issues in key micro-areas, poor stress uniformity on the sealing surface, and increased local wear, reducing valve sealing life and long-term operational stability. Furthermore, adapting to new operating conditions requires multiple physical prototype trials, resulting in long process optimization cycles and low iteration efficiency, making it difficult to adapt to the rapidly evolving needs of the semiconductor equipment industry. Summary of the Invention

[0004] This invention provides a method and system for optimizing the honing trajectory of semiconductor valve sealing surfaces based on spatial attention mechanism. It achieves adaptive and precise optimization of the honing trajectory, improves the processing quality of the sealing surface and the valve sealing performance, shortens the process optimization cycle, and adapts to the industry demand for rapid updates of semiconductor equipment.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect is a method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism, the method comprising: Construct a database linking sealing surfaces and processes; Based on the sealing surface-process correlation database, a pre-set convolutional neural network model is trained to obtain the trained convolutional neural network model. The morphological data of the sealing surface to be processed is input into the trained convolutional neural network model, and the key micro-regions on the sealing surface are located through its spatial attention mechanism; multiple morphological reference points corresponding to the key micro-regions are extracted from the sealing surface-process association database to form a morphological reference point dataset. Based on the spatial coordinates of each point in the topographic reference point dataset, a convex hull polyhedron and a minimum volume covering ellipsoid are constructed respectively. The volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid is calculated and defined as the topographic envelope coefficient. The honing trajectory parameters are optimized using preset sealing performance indicators and topographic envelope coefficients to obtain optimized honing trajectory parameters. Based on the optimized honing trajectory parameters, the operating state of the sealing surface under various working conditions is simulated, and the internal leakage, wear and stress uniformity indicators of helium detection are verified. If the verification results do not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weight, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, and the final honing trajectory parameters are obtained. Based on the final honing trajectory parameters, the corresponding honing process scheme is generated.

[0006] Furthermore, a sealing surface-process correlation database is constructed, including: Collect morphological feature data, honing process parameters and corresponding sealing performance indicators of the sealing surface of semiconductor valves to obtain a raw dataset containing multi-dimensional feature parameters; Normalization preprocessing is performed on each parameter in the original dataset containing multidimensional feature parameters to eliminate numerical differences between different units and obtain a normalized standard dataset. The normalized standard dataset is stored according to a preset data structure to form a sealing surface-process related database.

[0007] Furthermore, based on the sealing surface-process correlation database, a pre-defined convolutional neural network model is trained to obtain a trained convolutional neural network model, including: The basic architecture for constructing a convolutional neural network model that includes a spatial attention module is established, whereby the spatial attention module is used to implement the spatial attention mechanism, resulting in an initialized convolutional neural network model. The data format of the constructed sealing surface-process association database is adjusted to match the data structure of the input layer of the initialized convolutional neural network model, forming an adapted input dataset; The adapted input dataset is loaded into the initialized convolutional neural network model to set the hyperparameters for model training, start the model training process, and obtain the trained convolutional neural network model.

[0008] Furthermore, the morphological data of the sealing surface to be processed is input into the trained convolutional neural network model, and the key micro-regions on the sealing surface are located through its spatial attention mechanism; multiple morphological reference points corresponding to the key micro-regions are extracted from the sealing surface-process correlation database to form a morphological reference point dataset, including: The topographic data of the sealing surface to be processed is input into the trained convolutional neural network model. The spatial attention mechanism is used to calculate the weights of the input topographic data in the spatial dimension, and an attention weight map containing the importance of each spatial position is obtained. The spatial locations with weight values ​​higher than a preset threshold in the attention weight map are analyzed and identified to determine the key micro-regions on the sealing surface that have the highest weight influencing sealing performance. The key micro-regions include the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip. The location index information of the key micro-regions in the sealing surface-process association database is obtained. Based on the location index information of the key micro-region, the coordinate data of the morphological reference point at the corresponding location is extracted from the sealing surface-process association database, and the extracted coordinate data of multiple morphological reference points are combined to form a morphological reference point dataset.

[0009] Furthermore, based on the spatial coordinates of each point in the topographic reference point dataset, a convex hull polyhedron and a minimum volume covering ellipsoid are constructed respectively. The volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid is calculated and defined as the topographic envelope coefficient. The honing trajectory parameters are optimized using preset sealing performance indicators and the topographic envelope coefficient to obtain the optimized honing trajectory parameters, including: Obtain the topographic reference point dataset, and construct the convex hull polyhedron and the minimum volume covering ellipsoid that enclose the topographic reference point dataset based on the spatial coordinates of each point in the dataset. Calculate the volume of the convex hull polyhedron and the volume of the minimum volume covering ellipsoid. Find the ratio of the volume of the convex hull polyhedron to the volume of the minimum volume covering ellipsoid, and define the ratio as the morphology envelope coefficient. The morphological envelope coefficient and the preset sealing performance index are jointly constructed into a multi-objective optimization function. The honing trajectory parameters are iteratively optimized and solved with the minimization of the multi-objective optimization function value as the optimization objective. The honing trajectory parameters that satisfy the preset convergence condition of the multi-objective optimization function value are obtained as the optimized honing trajectory parameters.

[0010] Furthermore, based on the optimized honing trajectory parameters, the operating state of the sealing surface under various working conditions is simulated, and the helium leak rate, wear rate, and stress uniformity indicators are verified. If the verification results do not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weights, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, resulting in the final honing trajectory parameters, including: The optimized honing trajectory parameters are input into a pre-built multi-condition dynamic simulation model, which includes operating condition parameters under rated conditions and extreme conditions. The multi-condition dynamic simulation model is used to simulate the operating state of the sealing surface under various conditions, and the simulation values ​​of helium internal leakage, wear, and stress uniformity under the corresponding conditions are obtained. The simulated values ​​are compared with preset thresholds for helium internal leakage, wear, and stress uniformity to obtain the verification results for each index. If the verification results show that any simulation value does not meet the corresponding preset threshold, deviation information is generated based on the deviation between the unmet item and the threshold, and fed back to the trained convolutional neural network model to adjust its spatial attention weights, triggering the re-execution of steps 3 to 4 until all simulation values ​​meet the corresponding preset thresholds, and the honing trajectory parameters of the current iteration cycle are used as the final honing trajectory parameters.

[0011] Furthermore, based on the final honing trajectory parameters, a corresponding honing process scheme is generated, including: Obtain the final honing trajectory parameters as the core process parameters to be output; Extract the matching process parameters corresponding to the final honing trajectory parameters from the sealing surface-process association database. The matching process parameters include at least honing pressure, rotational speed and feed rate. Collect the simulated values ​​of helium internal leakage, wear, and stress uniformity output by the multi-condition dynamic simulation model under various working conditions, as well as the verification results of all simulation values ​​meeting the preset threshold, and integrate them to form a performance verification report. The final honing trajectory parameters, their corresponding process parameters, and the generated performance verification report are assembled according to the set process file format to generate a complete honing process scheme containing core process parameters, supporting process parameters, and a performance verification report.

[0012] Secondly, a semiconductor valve sealing surface honing trajectory optimization system based on spatial attention mechanism includes: The database construction module is used to collect morphological feature data of the sealing surface of semiconductor valves, honing process parameters and corresponding sealing performance indicators, perform normalization preprocessing, and construct a sealing surface-process association database. The model building and training module is used to train a preset convolutional neural network model based on the sealing surface-process association database to obtain the trained convolutional neural network model. The key micro-area localization and dataset generation module is used to input the morphology data of the sealing surface to be processed into the trained convolutional neural network model, and locate the key micro-area on the sealing surface through its spatial attention mechanism; extract multiple morphology reference points corresponding to the key micro-area from the sealing surface-process association database to form a morphology reference point dataset. The trajectory parameter optimization and solution module is used to construct a convex hull polyhedron and a minimum volume covering ellipsoid based on the spatial coordinates of each point in the topography reference point dataset. It calculates the volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid and defines it as the topography envelope coefficient. The honing trajectory parameters are optimized using preset sealing performance indicators and topography envelope coefficients to obtain the optimized honing trajectory parameters. The multi-condition simulation verification iteration module is used to simulate the operating state of the sealing surface under various conditions based on the optimized honing trajectory parameters, and to verify the helium detection internal leakage, wear amount and stress uniformity index. If the verification result does not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weight, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, and the final honing trajectory parameters are obtained. Based on the final honing trajectory parameters, the corresponding honing process scheme is generated.

[0013] The above-described solution of the present invention has at least the following beneficial effects: By employing a technique that constructs a database linking the sealing surface and the process, and builds a convolutional neural network model with a spatial attention mechanism, the technical problems of traditional honing trajectory optimization—such as lack of feature focusing capability and failure to perform morphological quantification analysis on key micro-areas—are overcome. This enables precise positioning and morphological feature quantification of key micro-areas on the sealing surface, improving the targeting and accuracy of honing trajectory optimization. Furthermore, by using a technique that calculates the morphological envelope coefficient and combines it with sealing performance indicators to construct a multi-objective optimization function, the technical problem of traditional processes relying solely on macroscopic performance indicators for optimization is overcome. This enables precise optimization of honing trajectory parameters, effectively avoiding over- or under-polishing of key micro-areas, improving the stress uniformity of the sealing surface, and reducing localized wear. Finally, by employing a technique that uses multi-condition dynamic simulation verification and constructs a closed-loop optimization system with deviation feedback iteration, the technical problem of traditional processes relying solely on static detection under a single condition and lacking closed-loop optimization capabilities is overcome. This allows for the prediction of the stability of the sealing surface under complex operating conditions, reasonably reducing the number of physical prototype trials, shortening the process optimization cycle, and adapting to the rapidly evolving needs of the semiconductor equipment industry. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of a method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism, provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of a semiconductor valve sealing surface honing trajectory optimization system based on spatial attention mechanism provided in an embodiment of the present invention. Detailed Implementation

[0016] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0017] like Figure 1 As shown, embodiments of the present invention propose a method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism. The method includes the following steps: Step 1: Construct a database linking sealing surfaces and processes; Step 2: Based on the sealing surface-process association database, train the preset convolutional neural network model to obtain the trained convolutional neural network model; Step 3: Input the morphology data of the sealing surface to be processed into the trained convolutional neural network model, and locate the key micro-regions on the sealing surface through its spatial attention mechanism; extract multiple morphology reference points corresponding to the key micro-regions from the sealing surface-process association database to form a morphology reference point dataset. Step 4: Based on the spatial coordinates of each point in the topographic reference point dataset, construct a convex hull polyhedron and a minimum volume covering ellipsoid respectively, calculate the volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid and define it as the topographic envelope coefficient; optimize the honing trajectory parameters with preset sealing performance indicators and topographic envelope coefficient to obtain the optimized honing trajectory parameters. Step 5: Based on the optimized honing trajectory parameters, simulate the operating state of the sealing surface under various working conditions, and verify the helium detection internal leakage, wear amount and stress uniformity index; if the verification result does not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weight, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, and the final honing trajectory parameters are obtained. Step 6: Based on the final honing trajectory parameters, generate the corresponding honing process scheme.

[0018] In this embodiment of the invention, by constructing a sealing surface-process correlation database and building a convolutional neural network model with a spatial attention mechanism, key micro-regions such as the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip can be accurately located, enabling quantitative analysis of the morphological features of key areas. Based on the morphological reference point dataset, a convex hull polyhedron and a minimum volume covering ellipsoid are constructed, the morphological envelope coefficient is calculated, and the honing trajectory parameters are optimized in conjunction with sealing performance indicators. This effectively avoids over-polishing or under-polishing problems in key micro-regions, improves the stress uniformity of the sealing surface, and reduces local wear. Through multi-condition dynamic simulation verification and the establishment of a closed-loop optimization system with deviation feedback iteration, the operational stability of the sealing surface under complex working conditions can be predicted, reasonably reducing the number of physical prototype trials and errors and shortening the process optimization cycle. The final complete honing process scheme can be directly used to guide actual processing, adapting to the industry demand for rapid updates in semiconductor equipment, and effectively improving the processing quality and sealing performance of valve sealing surfaces.

[0019] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Collect morphological feature data, honing process parameters, and corresponding sealing performance indicators of the semiconductor valve sealing surface to obtain a raw dataset containing multi-dimensional feature parameters. Specifically, this includes: collecting full-dimensional feature parameters of the sealing surface of semiconductor valves used in core semiconductor processes such as etching, deposition, and ion implantation before, during, and after processing; collecting morphological feature data involves scanning key micro-regions on the sealing surface that play a decisive role in sealing performance point by point, obtaining the three-dimensional spatial coordinates and surface contour parameters of each key micro-region, and recording the spatial position information and morphological feature values ​​of each measurement point; the key micro-regions include at least the root of the sealing surface cutting edge, the intersection of the transition arc and the sealing plane. The sealing surface includes the boundary and the inner and outer edges of the sealing strip. The root of the sealing surface edge refers to the acute-angle transition area formed by the edge of the sealing surface and the valve stem axis. This area is a core sensitive point for stress concentration and sealing pressure distribution, and its morphological accuracy directly affects the initial fit quality and long-term fatigue resistance of the sealing surface. The junction between the transition arc and the sealing plane refers to the boundary position where the arc transition area smoothly connects to the planar sealing surface. The continuity of the morphology in this area directly determines the uniformity of the sealing strip and the contact pressure distribution. The inner and outer edges of the sealing strip refer to the two boundary contour lines of the actual contact area of ​​the sealing surface. The inner edge determines the effective pressure bearing range of the sealing surface, while the outer edge affects the machining allowance and assembly accuracy of the sealing surface.

[0020] The acquisition of honing process parameters involves reading processing parameters such as honing pressure, spindle speed, feed rate, abrasive type, and abrasive particle size from the honing process execution record, ensuring that each set of morphological feature data has a corresponding complete process parameter record. The acquisition of sealing performance indicators involves obtaining data on helium leak detection, wear, and stress uniformity of the sealing surface through a standardized testing process after the corresponding process parameters are completed. Helium leak detection is used to characterize the leakage control capability of the sealing surface, wear is used to characterize the wear resistance of the sealing surface, and stress uniformity is used to characterize the stress distribution state across the entire sealing surface. All collected morphological feature data, honing process parameters, and sealing performance indicators are categorized, organized, and summarized according to valve model, process batch, and testing conditions. A many-to-one correlation is established between morphological features and process parameters, and between process parameters and performance indicators, generating a multi-dimensional feature parameter raw dataset containing three major categories: morphological features, process parameters, and performance indicators.

[0021] Step 1.2 involves normalizing the parameters in the original dataset containing multidimensional feature parameters to eliminate numerical differences between different units, resulting in a normalized standard dataset. Specifically, this includes: firstly, classifying and verifying the morphological feature data, honing process parameters, and sealing performance indicators in the original dataset; removing invalid data due to acquisition or recording errors; filling in a small amount of missing data due to acquisition omissions; and identifying and removing abnormal data exceeding reasonable limits to ensure the completeness and accuracy of the dataset. Secondly, addressing the differences in units, dimensions, and numerical magnitudes among different types of parameters, a maximum-minimum method is used. The value normalization method performs linear numerical transformation on each effective parameter. That is, for each type of parameter, the maximum and minimum values ​​are calculated to determine the numerical fluctuation range of the parameter. Then, based on the fluctuation range, the original value of each parameter is linearly scaled to a uniform standard interval, effectively eliminating numerical differences between different units and avoiding weight bias in subsequent model training due to different parameter numerical magnitudes. After processing, the transformation results of all parameters are checked a second time to confirm that the numerical transformation is without bias and the interval mapping is without error. Finally, a normalized standard dataset with accurate data, dimension matching, and can be directly used for model training is generated.

[0022] Step 1.3 involves storing the normalized standard dataset according to a pre-defined data structure to construct a sealing surface-process related database. Specifically, this includes: considering the actual business requirements for optimizing the honing trajectory of semiconductor valve sealing surfaces, matching the input data format, parameter extraction logic, and data retrieval efficiency requirements of the next-stage convolutional neural network model, while also taking into account the database's scalability and data traceability; and pre-designing a standardized database data structure with a three-level hierarchy, traceable associations, and support for multi-dimensional retrieval. The basic level is completed according to sealing surface basic attribute categories, honing process categories, and sealing performance categories. The system is categorized into three levels: First, the primary category is further subdivided by parameter characteristics. For example, under morphological feature data, key micro-area parameters and overall morphological parameters are subdivided; under honing process parameters, equipment operating parameters and abrasive characteristic parameters are subdivided; and under sealing performance indicators, core performance parameters and auxiliary performance parameters are subdivided. Second, each secondary sub-category is refined according to parameter acquisition standards and testing conditions. Each data record is assigned a unique global identifier code, which includes characteristic information such as valve model, process batch, and testing conditions. This achieves a precise, multi-dimensional mapping of basic attributes, process parameters, and performance indicators.

[0023] The parameters in the normalized standard dataset are sequentially and accurately entered into the corresponding database storage nodes according to the classification requirements, association rules, and identification coding rules of the three-level hierarchical structure. During the entry process, data format verification and association relationship verification are completed simultaneously to avoid problems such as parameter mismatch and missing identification. At the same time, the database is configured with multi-dimensional search functions, supporting data retrieval and extraction by single or combined conditions such as valve model, process type, key parameter threshold, and detection conditions. A data update and version management module is built to standardize the version records of subsequently added sealing surface morphology data, honing process parameters, and sealing performance indicators, ensuring the dynamic scalability of the database. Finally, a sealing surface-process association database that is complete, structurally standardized, accurately associated, efficiently retrieved, scalable, and traceable is constructed.

[0024] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Construct the basic architecture of a convolutional neural network model including a spatial attention module. The spatial attention module implements the spatial attention mechanism to obtain the initialized convolutional neural network model. Specifically, this includes: designing an adapted convolutional neural network architecture based on the actual needs of optimizing the honing trajectory of the semiconductor valve sealing surface, combined with the feature extraction rules of multi-dimensional parameters in the sealing surface-process correlation database. The architecture is built sequentially according to the logical hierarchy of data input, shallow feature extraction, deep feature extraction, feature dimensionality reduction, feature integration, and parameter output. The overall architecture includes an input layer, multiple convolutional layers, multiple pooling layers, a fully connected layer, and an output layer. The input layer receives standardized sealing surface-related parameter data. The convolutional layers extract shallow local features and deep overall features by setting convolutional kernels of different sizes. The pooling layer works with the convolutional layers to achieve feature dimensionality reduction and retain effective feature information. The fully connected layer integrates and fits all extracted features. The output layer outputs the finally optimized honing trajectory parameters.

[0025] In the aforementioned infrastructure, a spatial attention module is precisely embedded between each convolutional layer and its corresponding pooling layer, enabling the spatial attention module to directly receive and process the sealing surface parameter feature map output by the convolutional layer. The spatial attention module first performs feature compression and global information aggregation on the input feature map in the channel dimension, and then performs point-by-point and region-by-region correlation calculation on the aggregated features in the spatial dimension. By judging the degree of influence of the corresponding position on the sealing performance through the degree of correlation between features, a spatial attention weight map with the same size as the original feature map is generated. At the same time, a weighting operation is performed on the original feature map, assigning higher weight values ​​to key micro-region features that significantly affect the sealing performance, such as the root of the sealing surface blade, the junction of the transition arc, and the inner and outer edges of the sealing strip, and assigning lower weight values ​​to features in non-key regions. This achieves the enhancement of key micro-region features and the suppression of non-key region features. The feature map after weight allocation is then transmitted to the adjacent pooling layer for subsequent feature dimensionality reduction processing, ultimately constructing a convolutional neural network initial model architecture with a complete structure, clear hierarchy, smooth functional connection, and spatial attention feature focusing capability.

[0026] Step 2.2 involves adjusting the data format in the constructed sealing surface-process association database to match the input structure of the initialized convolutional neural network model, forming an adapted input dataset. Specifically, this includes: performing a comprehensive input layer analysis on the completed initial convolutional neural network model architecture containing spatial attention mechanisms; confirming the core requirements for the input layer, such as the acceptable numerical data dimensions, data storage types, data arrangement order, batch input data volume, and data arrangement format, to form a standardized input data specification suitable for this model training; and then uniformly retrieving the normalized standard dataset from the sealing surface-process association database and processing each element in the dataset according to the aforementioned standardized input data specification. The parameters undergo sequential format conversion and structural reorganization. Morphological feature data, honing process parameters, and sealing performance indicators are systematically integrated according to the specified dimensions required by the model. The originally discrete parameters are uniformly arranged in a preset fixed order. Throughout the conversion and reorganization process, the correlation mapping relationship and unique identification code between various parameters are strictly preserved to avoid problems such as parameter mismatch, correlation breakage, and identification loss during data adjustment. After the data adjustment is completed, the new dataset is subjected to full input adaptability verification. The data dimensions, arrangement order, and data format are checked item by item to ensure that they fully match the requirements of the input layer of the initial model architecture. After confirmation, the finally adapted input dataset that can be directly and stably loaded into the model is generated.

[0027] Step 2.3 involves loading the adapted input dataset into the initialized convolutional neural network model to set the hyperparameters for model training, starting the model training process, and obtaining the trained convolutional neural network model. Specifically, this includes: randomly dividing the adapted input dataset into training, validation, and test sets according to a preset ratio. The training set is used for fitting the core parameters of the model, the validation set is used for performance verification and parameter adjustment during model training, and the test set is reserved for final performance testing after model training is complete. Then, the training and validation sets are batch-divided according to the amount of input data per batch and fully loaded into the initial convolutional neural network model architecture containing a spatial attention mechanism. Based on the dual-objective requirement of optimizing the honing trajectory of the semiconductor valve sealing surface, reasonable hyperparameters are set, with the initial learning rate set to 0.001, the training batch size set to 32 or 64 based on the amount of input data, and the number of iterations set to 100 to 200. An early stopping mechanism is configured to prevent model overfitting. For the loss function, the mean squared error loss function is selected as the core basic loss function, considering the multi-objective optimization requirements. Its mathematical expression is as follows:

[0028] In the formula, m The total number of training samples, k The training sample number. For the firstk The actual optimal honing trajectory parameters corresponding to each training sample. Therefore, this model is for the first k The honing trajectory parameters predicted from each training sample. This is used to accurately measure the numerical deviation between the honing trajectory parameters predicted by the model and the actual optimal parameters; at the same time, a weighted fusion method is used to incorporate the morphology envelope coefficient and sealing performance index into the overall loss, constructing a dual-objective fusion loss function, the formula of which is as follows:

[0029] in, This is the total loss function ultimately used for model training. The mean squared error loss for trajectory parameter prediction. This is the loss term corresponding to the morphology envelope coefficient. This represents the loss item corresponding to the sealing performance index. The weighting coefficients are adjustable and satisfy the following conditions: In the context of semiconductor precision sealing processing, the weighting coefficient is set to... .

[0030] The specific fusion process is as follows: First, the trajectory parameter error loss, morphology regularity loss, and sealing performance loss are calculated separately, and then linearly superimposed according to the above weight ratios. During training, the weight coefficients are dynamically fine-tuned according to the priority requirements of different semiconductor process scenarios to achieve precise adaptation of dual-objective optimization. For the optimizer, an adaptive moment estimation optimizer is selected. This optimizer can dynamically adjust the learning rate according to the gradient changes of different dimension parameters during model training, taking into account both the training efficiency and convergence stability of the model, and adapting to the training characteristics of multi-dimensional and multi-type parameters in the sealing surface-process correlation database. After completing the configuration of all hyperparameters, loss functions, and optimizers, the formal training process of the convolutional neural network model containing spatial attention mechanism is started, allowing the model to continuously fit the parameter features during training, optimize the prediction accuracy of honing trajectory parameters, and finally obtain the trained convolutional neural network model.

[0031] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Input the topographic data of the sealing surface to be processed into the trained convolutional neural network model. Calculate the spatial weights of the input topographic data using its spatial attention mechanism to obtain an attention weight map containing the importance of each spatial location. Specifically, this includes: acquiring complete topographic data of the sealing surface of the semiconductor valve to be processed, which includes the three-dimensional spatial coordinates and surface contour parameters of the entire sealing surface; standardizing the topographic data according to the input format required by the trained convolutional neural network model to generate a topographic feature vector matching the model's input layer; inputting the topographic feature vector into the trained convolutional neural network model, which sequentially passes the data to the input layer, convolutional layer, spatial attention module, pooling layer, and fully connected layer according to the forward propagation logic; when the topographic feature map is transmitted to the spatial attention module, the spatial attention module performs global feature aggregation on the feature map in the spatial dimension, calculating the degree of difference of each spatial location on the feature map relative to the overall features; specifically, assuming the feature map has a total of... N The spatial location, the first p The feature vector at each position is denoted as . , No. The feature vector at each position is denoted as . ; Arithmetic mean of all positional eigenvectors As a global aggregation feature, then the first p Global dissimilarity at each location Calculate using the following formula: In the formula, This indicates the absolute value operation. Used to measure the p The degree to which the characteristics of a location deviate from the overall average; the larger the value, the higher the potential impact of that spatial location on sealing performance.

[0032] Simultaneously, to eliminate computational fluctuations caused by measurement noise or local anomalies in single-point features, feature information from a preset neighborhood surrounding the current spatial location is extracted; let the first... p The set of all locations within the neighborhood of a given location is: The number of locations within the neighborhood is Then the mean of local regional features This is used to characterize the overall features of the area surrounding the location; the single-point deviation value is combined with the mean of the neighborhood features to obtain the first... p Base score of feature importance at each location The formula is as follows: ; After calculating the base scores for all spatial locations, the base scores are normalized, i.e., the sum of the base scores for all locations is calculated. And divide the base score of each position by this sum to get the first position.p Standardized weight values ​​for each position Therefore, the weight value for each spatial location All are mapped to [0 , Within the interval [1], and the sum of all weight values ​​is 1; in this weight allocation process, for regions that significantly contribute to sealing performance, such as the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip, higher weight values ​​are assigned because their features deviate more from the overall average and their neighborhood feature averages are larger; for regions on the sealing surface that have a weak impact on sealing performance, lower weight values ​​are assigned because their features are closer to the overall average and their neighborhood fluctuations are smaller; finally, an attention weight map with the same size as the input morphology feature map and which can completely reflect the distribution of the importance of features at each spatial location on the sealing surface is generated.

[0033] Step 3.2: Analyze and identify spatial locations in the attention weight map where the weight value is higher than a preset threshold, and determine the key micro-regions on the sealing surface that have the highest weight influencing sealing performance. The key micro-regions include the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip. Obtain the position index information of the key micro-regions in the sealing surface-process association database. Specifically, this includes: based on the actual precision requirements and sealing performance indicators of the honing process of the semiconductor valve sealing surface, combined with historical process data and the structural characteristics of the sealing surface, a weight screening threshold is preset. The setting of the weight screening threshold is not a single fixed value, but is dynamically configured according to the priority of different process scenarios. The specific setting process is as follows: first, for different application scenarios of semiconductor valves, such as high-purity gas transportation scenarios for etching equipment, process gas control scenarios for deposition equipment, and ultra-high vacuum sealing scenarios for ion implantation equipment, the attention weight distribution characteristics of key and non-key micro-regions of the sealing surface in the historical processing data of each scenario are statistically analyzed, and the correspondence between scenarios and weight thresholds is established.

[0034] By combining the sealing surface structure drawings with industry standards, the theoretical weight ranges for the root of the sealing surface edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip under different working conditions were determined. The theoretical weight ranges were cross-validated with the actual statistical distribution to obtain the optimal boundary value for distinguishing between critical and non-critical areas in each scenario. Based on the tolerance requirements of semiconductor valve sealing surface processing, a reasonable floating range was set for the weight screening threshold. In the high-purity gas transportation scenario with extremely high sealing requirements, the weight screening threshold was set in a higher range to ensure accurate identification and focus on all small areas that significantly affect helium internal leakage detection and stress uniformity. In the high-precision process chamber pressure control scenario that emphasizes morphological accuracy control, the threshold was set in a relatively mild range to ensure that the model has a comprehensive perception of the overall morphological features.

[0035] After setting the threshold, the weight value of each spatial location in the generated attention weight map is compared with the preset threshold one by one. Spatial locations with weight values ​​higher than the preset threshold are marked as high-influence feature regions, and spatial locations with weight values ​​lower than the preset threshold are marked as low-influence regions. Combining the actual structural distribution and process characteristics of the sealing surface, key micro-regions that play a crucial role in sealing performance are accurately identified from the high-influence feature regions. The key micro-regions include the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip. For each of the above key micro-regions, its two-dimensional coordinate origin, coordinate extension range, and region boundary contour on the sealing surface are extracted. The center coordinates of each key micro-region are used as the reference index point, and the coordinate extension range of the micro-region relative to the overall coordinate system of the sealing surface is recorded. At the same time, the feature contour code of the micro-region boundary is defined. By orderly combining the reference index point coordinates, extension range parameters, and boundary contour codes, a set of structured, unique, and quickly searchable location index information is generated. This location index information can be directly used as a retrieval pointer to locate and extract the morphological reference point coordinate data of the corresponding position in the sealing surface-process association database.

[0036] Step 3.3: Based on the location index information of the key micro-regions, extract the coordinate data of the morphological reference points at the corresponding locations from the sealing surface-process association database. Combine the extracted coordinate data of multiple morphological reference points to form a morphological reference point dataset. Specifically, this includes: using the unique location index information corresponding to each key micro-region as the basis for precise retrieval, performing structured positioning and step-by-step data backtracking operations in the sealing surface-process association database; matching the reference coordinates, region extension range, and boundary contour codes contained in the index information with the pre-stored sealing surface geometric dimension parameters, morphological measurement point coordinates, and region classification identifiers in the database to quickly locate the storage address and target data entry corresponding to each key micro-region in the database; and extracting the key micro-regions from the corresponding storage nodes in the database according to the retrieval rules pointed to by the index information. The coordinate data of the topographic reference points corresponding to the root of the cover cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip are provided. Each topographic reference point coordinate data contains the precise position information of the point in three-dimensional space. Multiple sets of topographic reference point coordinates extracted from different key micro-regions are centrally organized and arranged in an orderly manner according to the type of key micro-region and the actual spatial distribution order of the sealing surface. At the same time, multi-dimensional validity verification is carried out on all extracted coordinate data, duplicate coordinate data is removed, missing coordinate entries are filled in, and abnormal coordinate data that exceeds the physical size and processing tolerance range of the sealing surface are filtered out. This ensures that each set of topographic reference point coordinates meets the requirements of the real processing scenario and model input, and guarantees the accuracy, completeness and consistency of the dataset. Finally, a topographic reference point dataset with a unified format, standardized fields, and usable for subsequent honing trajectory optimization calculations is formed.

[0037] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Obtain the topographic reference point dataset. Based on the spatial coordinates of each point in the topographic reference point dataset, construct the convex hull polyhedron and the minimum volume covering ellipsoid that enclose the topographic reference point dataset. Specifically, this includes: obtaining the topographic reference point dataset generated in Step 3.3 after repeatability, integrity, and rationality checks. This dataset consists of three-dimensional spatial coordinate points of three key micro-regions: the root of the sealing surface cutting edge, the intersection of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip. Each coordinate point contains standardized and unique X-axis, Y-axis, and Z-axis position information. All coordinate points together reflect the true three-dimensional topographic distribution characteristics of the key micro-regions of the sealing surface. Based on the three-dimensional spatial coordinates of all topographic reference points in the dataset, use an incremental convex hull algorithm to construct a convex hull polyhedron that can completely enclose all reference points point by point.

[0038] The specific construction process is as follows: Starting from any three non-collinear reference points, an initial tetrahedral convex hull structure is constructed. Then, each remaining reference point is added to the current convex hull structure in turn. By analyzing the spatial positional relationship between the point and each outer surface of the convex hull, it is determined whether the point is outside the convex hull. If the point is outside the convex hull, all outer surfaces of the convex hull that can be observed from the point are retained, and the inner hidden surfaces that cannot be observed are deleted. The point is then connected to the boundary edges of the retained surfaces to form a new set of outer surfaces, thereby completing the dynamic update of the convex hull vertices, edges, and sides. The operations of adding points, determining positions, updating sides, and reconstructing structures are repeated until all topographic reference points are included in the convex hull structure. Finally, a convex hull polyhedron with no depressions, no overlaps, no redundancy, and the smallest volume is obtained. This convex hull polyhedron is composed of several planar triangular sides, and all sides maintain an outward convex shape, forming a compact and closed three-dimensional enveloping space.

[0039] Simultaneously, based on the same set of topographic reference point datasets, an iterative calculation of the minimum volume ellipsoid covering all reference points is performed using a method based on the covariance matrix. Specifically, the process involves: calculating the arithmetic mean of the three-dimensional coordinates of all topographic reference points to obtain the geometric center coordinates of the dataset; then centering all reference point coordinates with the geometric center as the origin; constructing the covariance matrix of the three-dimensional coordinate data based on the centered coordinates; and performing eigenvalue decomposition on the covariance matrix. The specific calculation process is as follows: The three-dimensional covariance matrix is ​​denoted as... By solving the characteristic equation Obtain the corresponding eigenvalues ,in It is a third-order identity matrix. Represents the determinant operation of a matrix. Let each eigenvalue be an eigenvalue to be solved. After obtaining the eigenvalues, substitute each eigenvalue into the system of linear equations. Solving the equations yields eigenvectors that correspond one-to-one with each eigenvalue. .

[0040] The three eigenvectors obtained are mutually orthogonal in space and are used to characterize the three main distribution directions of the topographic reference points in three-dimensional space. After the eigenvalues ​​are square-rooted and normalized, the semi-axis lengths of the minimum volume covering ellipsoid in the three principal axis directions are obtained. Combining the geometric center coordinates, the three principal axis directions and the corresponding semi-axis lengths, the minimum volume covering ellipsoid that can completely cover all topographic reference points and has the smallest volume among all feasible ellipsoids is uniquely determined.

[0041] Step 4.2: Calculate the volume of the convex hull polyhedron and the volume of the minimum volume covering ellipsoid. Calculate the ratio of the volume of the convex hull polyhedron to the volume of the minimum volume covering ellipsoid, defining this ratio as the topographic envelope coefficient. Specifically, after constructing the convex hull polyhedron and the minimum volume covering ellipsoid, the convex hull polyhedron undergoes rigorous spatial partitioning. Using the centroid of the convex hull polyhedron as a common vertex, each outer surface of the convex hull polyhedron is combined with this common vertex to form several non-overlapping tetrahedral elements whose vertices are all topographic reference points and completely fill the convex hull space. Then, the volume of each tetrahedral element is calculated using the vector determinant method. The formula for calculating the volume of a single tetrahedron is: ,in Let be the three-dimensional spatial vectors corresponding to the three edges of a tetrahedron originating from the same vertex, and × represent the vector cross product operation. This refers to the dot product operation of vectors. For absolute value operations, after obtaining the volumes of all tetrahedral elements, the volumes of all tetrahedral elements are summed to obtain the total volume of the convex hull polyhedron. ,in For the first i The volume of a tetrahedral unit cell n This represents the total number of tetrahedral elements obtained after partitioning the convex hull polyhedron. From the 1st to the 2nd n Volume accumulation operation of tetrahedral elements.

[0042] Simultaneously, based on the constructed minimum volume covering ellipsoid, its center position in the three-dimensional coordinate system and the lengths of its three mutually perpendicular principal semi-axes are determined. The volume of the minimum volume covering ellipsoid is then calculated using the standard geometric volume formula. Where π is the constant of pi. a, b, cLet these be the lengths of the first, second, and third principal semi-axes, which are perpendicular to each other in three-dimensional space, representing the minimum volume covering ellipsoid. After accurately calculating the volumes of the convex hull polyhedron and the minimum volume covering ellipsoid, the ratio between them is calculated, and this ratio is formally defined as the shape envelope coefficient of the shape reference point set. The formula for calculating the shape envelope coefficient is as follows: ,in K The shape envelope coefficient, Let the volume of the convex hull polyhedron be . To minimize the volume of the ellipsoid, this morphological envelope coefficient can quantitatively characterize the compactness, geometric regularity, and spatial aggregation characteristics of the key micro-region morphological reference points of the sealing surface in three-dimensional space. The smaller the coefficient value, the more regular and compact the distribution of the morphological reference points; the larger the coefficient value, the more discrete and loose the distribution of the morphological reference points.

[0043] Step 4.3: The morphology envelope coefficient and the preset sealing performance index are jointly constructed into a multi-objective optimization function. The honing trajectory parameters are iteratively optimized by minimizing the value of the multi-objective optimization function, resulting in honing trajectory parameters that satisfy the preset convergence condition of the multi-objective optimization function. These are then used as the optimized honing trajectory parameters. Specifically, this includes: first, normalizing the calculated morphology envelope coefficient and the preset sealing performance index of the sealing surface to unify the two parameters with different dimensions into the same numerical range; then, combining the normalized morphology envelope coefficient and the sealing performance index using a linear weighting method to construct a multi-objective optimization function for model iterative optimization. The mathematical expression of the multi-objective optimization function is: ;in The overall objective value of the multi-objective optimization function. morphological envelope coefficient K It is used to characterize the spatial regularity of the benchmark points for key micro-region morphology. These are dimensionless, normalized values ​​of sealing performance indicators, used to reflect the sealing capacity, pressure resistance rating, and leakage control level of the sealing surface. These are the weighting coefficients corresponding to the morphology envelope coefficients. These are the weighting coefficients corresponding to the sealing performance indicators, and the weighting coefficients satisfy the constraint conditions. It can be dynamically adjusted according to the emphasis on morphological accuracy and sealing performance in the actual processing scenario.

[0044] With multi-objective optimization function Minimizing the value of is the overall optimization objective. A complete backpropagation training process is performed on the convolutional neural network model with embedded spatial attention modules. During training, the gradient information of the optimization function with respect to the parameters of each layer of the model is calculated layer by layer based on the chain rule. The chain rule states that for a multi-layered nested composite function, its overall gradient is equal to the product of the gradients of each layer's function. In this model, this is manifested by continuously multiplying the gradient of the multi-objective optimization function with respect to the final output, the gradient of the output with respect to the intermediate feature maps, and the gradient of the intermediate feature maps with respect to the parameters of each layer, thereby obtaining the gradient of the optimization function with respect to the parameters of the lowest layer of the model. This achieves the layer-by-layer transfer of gradient information from the output layer to the input layer. The corresponding gradient calculation expression is: ;in θ This represents all parameters in the model that need to be iteratively updated, including the convolutional kernel weights of convolutional layers, the bias values ​​corresponding to the activation functions of each layer, and the weights and bias parameters of fully connected layers. These are the honing trajectory parameters output by the final forward propagation of the model. This represents the gradient of the multi-objective optimization function with respect to the model's output trajectory parameters. This represents the gradient of the model output with respect to the parameters of each internal layer.

[0045] Based on the calculated gradient information, the adaptive gradient descent optimization algorithm is used to update the model parameters round by round. The parameter update rule expression is as follows: ;in To complete the updated model parameters in this round, These are the model parameters before this update. η A pre-set learning rate is used to control the step size of parameter updates, avoiding the model failing to converge due to an excessively large update step size or the training efficiency being low due to an excessively small step size. After each round of parameter updates, the topographic reference point dataset is re-inputted into the model to perform forward propagation, obtaining the honing trajectory parameters under the current parameters and recalculating the multi-objective optimization function value. The process of forward propagation, gradient calculation, parameter back-update, and target value verification is continuously executed in a loop until the multi-objective optimization function value is lower than the preset convergence threshold for several consecutive rounds (this preset convergence threshold is used to determine whether the model training has reached a stable optimization state that meets the processing requirements; its value is usually a very small positive number, which can ensure that the honing trajectory parameters output by the model meet the requirements of actual engineering use in terms of topographic regularity and sealing performance) or the function value no longer shows a significant decrease and tends to stabilize. At this time, it is determined that the model has reached the convergence state, and the honing trajectory parameters output by the current model are used as the optimized honing trajectory parameters.

[0046] In this embodiment, the synergistic relationship between the dual-objective fusion loss function and the multi-objective optimization function is explained as follows: In this invention, the dual-objective fusion loss function described in step 2.3 and the multi-objective optimization function described in step 4.3 are not the same concept. There is no logical conflict between them. Instead, they are a collaborative relationship with clear division of labor and consistent objectives in the complete optimization process, as detailed below: Distinguishing between concepts and stages of function: The dual-objective fusion loss function serves as a supervisory signal during the neural network training stage. It guides the model to fit honing trajectory parameters that balance morphological regularity and sealing performance by weighting and fusing trajectory parameter prediction errors, morphological envelope coefficient deviations, and sealing performance index deviations. It only takes effect during model training. The multi-objective optimization function serves as an evaluation criterion during the process verification and iterative optimization stages. Based on the dynamic simulation results under multiple operating conditions, it comprehensively evaluates indicators such as helium internal leakage, wear of surface material on the sealing surface, and uniformity of stress distribution across the sealing domain. It is used to screen optimized honing trajectory parameters that meet actual service requirements and takes effect after the model outputs candidate trajectory parameters.

[0047] Collaborative Working Mechanism: The core optimization directions of the dual-objective fusion loss function and the multi-objective optimization function are highly unified, both aiming to improve sealing performance and ensure morphological regularity, with no conflicting objectives. Similarly, their operation phases are completely separated, and they will not take effect simultaneously in the same process, avoiding ambiguity and conflict at the definition or calculation level. The dual-objective fusion loss function and the multi-objective optimization function form a closed-loop collaboration: that is, the dual-objective fusion loss function provides the model with high-quality candidate honing trajectory parameters, and the multi-objective optimization function performs final performance verification on the candidate parameters. If the verification fails, a deviation feedback signal is generated and fed back to the model, driving the model to optimize again guided by the dual-objective fusion loss function, jointly completing the iterative optimization of the honing trajectory parameters.

[0048] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Input the optimized honing trajectory parameters into the pre-constructed multi-condition dynamic simulation model. The multi-condition dynamic simulation model includes operating condition parameters under rated conditions and extreme conditions. Specifically, it includes: obtaining the optimized honing trajectory parameters obtained after iterative optimization by the multi-objective optimization function in step 4.3, and importing this set of parameters as core input data into the pre-constructed multi-condition dynamic simulation model. The multi-condition dynamic simulation model adopts a hybrid architecture that deeply couples a deep convolutional neural network with a finite element solver. It consists of five parts: a feature input layer, a finite element physical constraint layer, an attention-enhanced simulation layer, a multi-condition mapping layer, and a result output layer. The layers are sequentially connected and work together to form a complete simulation inference chain from data input, physical constraints, region enhancement, condition adaptation to result output.

[0049] The feature input layer is responsible for receiving and unifying heterogeneous data from multiple sources, such as sealing surface geometric parameters, material mechanical properties, honing trajectory coordinates, and working condition load boundaries. It performs alignment, completion, and format standardization on the data, converting it into a unified format that the model can directly read and calculate. The finite element physical constraint layer is used to load engineering physical constraints such as sealing surface contact judgment rules, material mechanical properties, assembly constraint relationships, and deformation boundary conditions. It sets simulation boundaries with reference to actual assembly and processing conditions, providing a real and reliable physical basis and numerical calculation foundation for the entire simulation calculation process. The attention enhancement simulation layer inherits and reuses the spatial attention mechanism used in the previous stage of this invention. Through automatic extraction and weight allocation of key area features, it automatically identifies and strengthens the simulation weight of key micro-areas that play a decisive role in sealing performance, such as the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip, thereby improving the calculation accuracy and detail reproduction capability of local areas.

[0050] The multi-condition mapping layer is used to automatically adapt to boundary conditions such as pressure fluctuations, temperature changes, load impacts, and operating time under different conditions. It completes parameter switching and state iteration according to the preset condition sequence, realizing dynamic switching of condition parameters and real-time updating of simulation state. The model integrates complete operating condition parameters corresponding to rated conditions and extreme conditions. Rated conditions include standard pressure, normal temperature, and rated medium transportation state when the equipment is working normally. Extreme conditions include harsh working conditions such as instantaneous overpressure, high and low temperature alternation, and long-term continuous operation. The result output layer summarizes, statistically analyzes, and quantifies the simulation calculation results, and uniformly outputs simulation results such as helium internal leakage, wear, and stress uniformity that can be directly used for index judgment.

[0051] Step 5.2: Simulate the operating state of the sealing surface under various working conditions using the multi-condition dynamic simulation model to obtain the simulated values ​​of helium internal leakage, wear, and stress uniformity under the corresponding working conditions. Specifically, this includes: starting and running the multi-condition dynamic simulation model according to the set multi-condition progressive simulation process. The model executes a complete simulation process in six consecutive stages: initialization, rated working condition simulation, extreme working condition simulation, data acquisition, index calculation, and result output. First, perform global initialization of the model, importing the optimized honing trajectory parameters, sealing surface geometry, material properties, and assembly parameters to be verified. The simulation steps, iteration count, calculation accuracy, and data sampling frequency are uniformly set. Then, the simulation enters the rated working condition simulation stage, applying standard working pressure, normal temperature, and stable medium transportation conditions. The load is applied slowly and evenly, and sufficient iterative calculations are completed to simulate the stress distribution of the sealing surface under normal working conditions, the contact fit between the valve seat and the valve core, and the entire process of material micro-deformation. Then, the simulation switches to extreme working conditions, applying harsh working conditions such as instantaneous overpressure, high and low temperature alternation, long-term continuous operation, and start-stop impact in sequence to fully reproduce the dynamic response and state change process of the sealing surface under harsh service environment.

[0052] Throughout the simulation process, information such as interface gaps, contact pressures, relative displacements, and stress distributions in key micro-areas, especially the root of the cutting edge, the junction of the transition arc, and the inner and outer edges of the sealing strip, are collected and stored in real time at fixed points, at fixed times, and with high density. After the simulation calculation is completed, the index calculation stage begins. When calculating the simulated value of helium detection internal leakage, based on the structural characteristics of the sealing surface and the gap distribution data collected in the simulation, gap parameters in key areas prone to leakage, such as the root of the cutting edge, the junction of the transition arc, and the inner and outer edges of the sealing strip, are selected. During the selection, measurement point data with larger gap values, continuous gap distribution, and the ability to form a through channel are prioritized, while invalid gap data that are isolated and have no connectivity trend are eliminated. Then, a comprehensive analysis is conducted in combination with the interface closure tightness and the working load conditions. By comparing the degree of opening and closing of the interface gap, the probability of channel connection, and the interface barrier capacity under different loads, the leakage risk is quantitatively assessed, and the simulated value of helium detection internal leakage is finally obtained.

[0053] Based on the collected data of contact pressure, relative sliding distance, and sliding frequency in various regions of the sealing surface, and combined with the material surface hardness and wear resistance characteristics, a point-by-point traversal and cumulative calculation is performed to obtain the simulated value of the wear amount of the surface material of the sealing surface. Based on the collected stress values ​​at various measuring points throughout the sealing area, after removing outliers, the distribution dispersion, regional fluctuation amplitude, and overall consistency of all values ​​are statistically analyzed to obtain the simulated value of the uniformity of stress distribution throughout the sealing area. Finally, the various simulation indicators under rated and extreme working conditions are classified, organized, and output uniformly to form a complete multi-condition simulation result file that can be directly used for subsequent verification and comparison.

[0054] Step 5.3: Compare the simulated values ​​with the preset thresholds for helium internal leakage, wear, and stress uniformity to obtain the verification results for each indicator. Specifically, this includes classifying and organizing the simulated values ​​for helium internal leakage, wear of the sealing surface material, and uniformity of stress distribution across the sealing area obtained under rated and extreme operating conditions. First, establish a fixed order for each operating condition and each indicator, based on the following: helium internal leakage under rated operating conditions, wear of the sealing surface material under rated operating conditions, uniformity of stress distribution across the sealing area under rated operating conditions, helium internal leakage under extreme operating conditions, wear of the sealing surface material under extreme operating conditions, and uniformity of stress distribution across the sealing area under extreme operating conditions. A unique and orderly one-to-one correspondence verification relationship is established to ensure that the internal leakage of helium under rated operating conditions matches only the internal leakage threshold of helium under rated operating conditions, the wear of the surface material of the sealing surface under rated operating conditions matches only the wear threshold of the corresponding rated operating conditions, the uniformity of stress distribution across the entire sealing area under rated operating conditions matches only the stress uniformity threshold of the corresponding rated operating conditions, the internal leakage of helium under extreme operating conditions matches only the internal leakage threshold of helium under extreme operating conditions, the wear of the surface material of the sealing surface under extreme operating conditions matches only the wear threshold of the corresponding extreme operating conditions, and the uniformity of stress distribution across the entire sealing area under extreme operating conditions matches only the stress uniformity threshold of the corresponding extreme operating conditions, thus eliminating cross-comparisons between different operating conditions and different indicators.

[0055] Simultaneously, each set of precisely matched simulation results is compared item by item with pre-set thresholds for helium leak, wear, and stress uniformity, based on semiconductor industry precision machining standards, valve sealing safety level requirements, honing process indicators, and actual service performance requirements. The comparison process strictly adheres to the principle of one-to-one comparison under the same operating conditions, for the same indicators, and within the same dimensions. Each set of simulation indicators is sequentially judged to determine whether it meets or exceeds the corresponding preset threshold requirements. Indicators that meet the preset requirements are marked as qualified, while those that do not are marked as requiring optimization. The operating condition type, deviation magnitude, and deviation direction corresponding to each indicator are recorded simultaneously. Finally, a complete and clear indicator verification result is generated, including the specific operating condition type, complete indicator name, actual simulation value, preset threshold upper limit, item-by-item comparison conclusion, qualified or requiring optimization status, and relevant deviation information.

[0056] Step 5.4: If the verification result shows that any simulation value does not reach the corresponding preset threshold, then deviation information is generated based on the deviation between the non-compliant item and the threshold, and fed back to the trained convolutional neural network model to adjust its spatial attention weights, triggering the re-execution of steps 3 to 4 until all simulation values ​​meet the corresponding preset thresholds. The honing trajectory parameters of the current iteration cycle are used as the final honing trajectory parameters. Specifically, if the verification result of the index shows that, under any of the rated or extreme working conditions, any one of the following is true: helium internal leakage, wear of the sealing surface material, or uniformity of stress distribution across the sealing area. If a simulation indicator fails to meet the corresponding preset threshold requirement, the non-compliant indicator is first precisely located and independently analyzed to identify the specific working condition, indicator type, simulation value, and corresponding threshold requirement. Then, the absolute value and direction of the deviation between the actual simulation value and the corresponding preset threshold are calculated to determine whether the deviation is too large or too small, and whether the degree of deviation is slight or severe. At the same time, combined with the working environment in which the deviation occurs and the distribution characteristics of key areas of the sealing surface, complete deviation information is formed, including the working condition to which the deviation belongs, the specific indicator name, the deviation value, the deviation direction, the degree of deviation, and the location of the area to be optimized.

[0057] Based on this deviation information, a deviation feedback signal with a precise optimization direction is generated. This feedback signal is then uniformly fed back into the trained convolutional neural network model, providing precise optimization guidance. The model is driven to eliminate the current deviation and meet all threshold requirements as optimization objectives, and the complete iterative optimization process of steps 3 to 4 is re-executed. This includes re-executing the location of key micro-regions and extraction of topographic reference points, calculation of topographic envelope coefficients and construction of multi-objective optimization functions, and iterative optimization of honing trajectory parameters. After obtaining new honing trajectory parameters in each round of iterative optimization, these parameters are again input into the multi-condition dynamic simulation model. In step 5, complete simulations, data acquisition, index calculations, and item-by-item comparisons and verifications are performed sequentially for both rated and extreme operating conditions, forming a closed-loop iterative mechanism of optimization, simulation, verification, and feedback. The iterative optimization and multi-condition verification process is repeated until the re-output honing trajectory parameters, after undergoing multi-condition dynamic simulation, meet their respective preset threshold requirements for the helium leak detection simulation value, the surface material wear value of the sealing surface, and the uniformity of stress distribution across the sealing domain under both rated and extreme operating conditions. At this point, the honing trajectory parameters output in the current iteration cycle and verified by all operating conditions and all indices are determined as the final honing trajectory parameters.

[0058] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1: Obtain the final honing trajectory parameters as the core process parameters to be output. Specifically, this includes: retrieving the final honing trajectory parameters that have been fully verified and qualified under rated and extreme conditions through the multi-condition dynamic simulation model in step 5.4, and whose simulation indicators meet the preset threshold requirements. Extract, confirm and solidify this set of parameters as the core process parameters of the entire honing process scheme, so as to provide stable and reliable basic data for the matching of subsequent supporting process parameters and the generation of the final process scheme.

[0059] Step 6.2: Extract the matching process parameters corresponding to the final honing trajectory parameters from the sealing surface-process association database. The matching process parameters include at least honing pressure, rotational speed, and feed rate. Specifically, the following steps are taken: using the determined final honing trajectory parameters as the retrieval basis, a hierarchical matching search is performed in the sealing surface-process association database. First, the trajectory type, trajectory density, and trajectory path distribution characteristics are used as the first-level matching conditions. Then, the sealing surface material properties, geometric structure dimensions, and machining accuracy level are used as the second-level matching conditions. The matching range is gradually narrowed, and the associated process record with the highest matching degree is selected. From this matching record, the matching process parameters that are highly compatible with the final honing trajectory parameters are extracted. Among them, honing pressure refers to the constant pressure applied to the surface of the sealing surface during the honing process, rotational speed refers to the speed at which the honing equipment spindle drives the tool to rotate, and feed rate refers to the speed at which the tool moves along the machining trajectory of the sealing surface. The extracted matching process parameters have a high degree of matching and good compatibility with the core honing trajectory parameters and can be directly applied to the actual honing process.

[0060] Step 6.3: Collect the simulated values ​​of helium internal leakage, wear, and stress uniformity output from the multi-condition dynamic simulation model under various operating conditions, as well as the verification results showing that all simulated values ​​meet the preset thresholds. Integrate these into a performance verification report. Specifically, this includes: firstly, extracting the simulation data corresponding to the rated operating condition, instantaneous overpressure operating condition, high and low temperature alternating operating condition, long-term continuous operation operating condition, and start-stop impact operating condition in sequence; then, independently analyzing the simulated values ​​of helium internal leakage, wear of the sealing surface material, and uniformity of stress distribution across the sealing area under each operating condition. The data is categorized and organized to ensure that data from different operating conditions and indicators are not mixed up. The results of each set of simulation values ​​and corresponding preset thresholds, the status indicators of qualified or unoptimized, and the information on the magnitude and direction of deviation are centrally summarized to form complete judgment information covering all operating conditions and all indicators. Then, the simulation values, threshold standards, comparison processes, and judgment conclusions of all operating conditions are arranged, logically sorted, and systematically integrated according to a unified format specification to finally form a performance verification report that is complete in data, well-founded, traceable in process, and can be used for technical review and document archiving.

[0061] Step 6.4 involves assembling the final honing trajectory parameters, their corresponding process parameters, and the generated performance verification report according to the set process document format. This generates a complete honing process plan containing core process parameters, supporting process parameters, and a performance verification report. Specifically, this includes: uniformly collecting the determined final honing trajectory parameters, extracted supporting process parameters, and generated performance verification report; standardizing them according to the process document standard structure and layout requirements stipulated in the semiconductor precision machining field; incorporating the final honing trajectory parameters into the core parameter page and the supporting process parameters into the supporting process page; incorporating the simulation data, comparison process, and pass / fail judgment results from the performance verification report into the simulation verification page and the quality judgment page; sequentially completing the classification, content filling, and format calibration according to the fixed structure; and then organically combining and systematically encapsulating all content to finally generate a standardized and complete honing process plan that is structurally complete, content-standardized, data-accurate, format-uniform, and can be directly used to guide actual honing processing execution, process control, quality inspection, and factory acceptance.

[0062] like Figure 2 As shown, embodiments of the present invention also provide a semiconductor valve sealing surface honing trajectory optimization system based on a spatial attention mechanism, comprising: The database construction module is used to collect morphological feature data of the sealing surface of semiconductor valves, honing process parameters and corresponding sealing performance indicators, perform normalization preprocessing, and construct a sealing surface-process association database. The model building and training module is used to train a preset convolutional neural network model based on the sealing surface-process association database to obtain the trained convolutional neural network model. The key micro-area localization and dataset generation module is used to input the morphology data of the sealing surface to be processed into the trained convolutional neural network model, and locate the key micro-area on the sealing surface through its spatial attention mechanism; extract multiple morphology reference points corresponding to the key micro-area from the sealing surface-process association database to form a morphology reference point dataset. The trajectory parameter optimization and solution module is used to construct a convex hull polyhedron and a minimum volume covering ellipsoid based on the spatial coordinates of each point in the topography reference point dataset. It calculates the volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid and defines it as the topography envelope coefficient. The honing trajectory parameters are optimized using preset sealing performance indicators and topography envelope coefficients to obtain the optimized honing trajectory parameters. The multi-condition simulation verification iteration module is used to simulate the operating state of the sealing surface under various conditions based on the optimized honing trajectory parameters, and to verify the helium detection internal leakage, wear amount and stress uniformity index. If the verification result does not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weight, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, and the final honing trajectory parameters are obtained. Based on the final honing trajectory parameters, the corresponding honing process scheme is generated.

[0063] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0064] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism, characterized in that, The method includes: Step 1: Construct a database linking sealing surfaces and processes; Step 2: Based on the sealing surface-process association database, train the preset convolutional neural network model to obtain the trained convolutional neural network model; Step 3: Input the morphology data of the sealing surface to be processed into the trained convolutional neural network model, and locate the key micro-regions on the sealing surface through its spatial attention mechanism; extract multiple morphology reference points corresponding to the key micro-regions from the sealing surface-process association database to form a morphology reference point dataset. Step 4: Based on the spatial coordinates of each point in the topographic reference point dataset, construct a convex hull polyhedron and a minimum volume covering ellipsoid respectively, calculate the volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid and define it as the topographic envelope coefficient; optimize the honing trajectory parameters with preset sealing performance indicators and topographic envelope coefficient to obtain the optimized honing trajectory parameters. Step 5: Based on the optimized honing trajectory parameters, simulate the operating state of the sealing surface under various working conditions, and verify the helium leak, wear, and stress uniformity indicators. If the verification results do not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weight, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, thus obtaining the final honing trajectory parameters.

2. The method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism according to claim 1, characterized in that, Construct a sealing surface-process association database, including: Collect morphological feature data, honing process parameters and corresponding sealing performance indicators of the sealing surface of semiconductor valves to obtain a raw dataset containing multi-dimensional feature parameters; Normalization preprocessing is performed on each parameter in the original dataset containing multidimensional feature parameters to eliminate numerical differences between different units and obtain a normalized standard dataset. The normalized standard dataset is stored according to a preset data structure to form a sealing surface-process related database.

3. The method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism according to claim 2, characterized in that, Based on the sealing surface-process correlation database, a pre-defined convolutional neural network model is trained to obtain the trained convolutional neural network model, including: The basic architecture for constructing a convolutional neural network model that includes a spatial attention module is established, whereby the spatial attention module is used to implement the spatial attention mechanism, resulting in an initialized convolutional neural network model. The data format of the constructed sealing surface-process association database is adjusted to match the data structure of the input layer of the initialized convolutional neural network model, forming an adapted input dataset; The adapted input dataset is loaded into the initialized convolutional neural network model to set the hyperparameters for model training, start the model training process, and obtain the trained convolutional neural network model.

4. The method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism according to claim 3, characterized in that, The topographic data of the sealing surface to be processed is input into the trained convolutional neural network model, and the key micro-regions on the sealing surface are located through its spatial attention mechanism. Multiple morphological reference points corresponding to the key micro-regions are extracted from the sealing surface-process correlation database to form a morphological reference point dataset, including: The topographic data of the sealing surface to be processed is input into the trained convolutional neural network model. The spatial attention mechanism is used to calculate the weights of the input topographic data in the spatial dimension, and an attention weight map containing the importance of each spatial position is obtained. The spatial locations with weight values ​​higher than a preset threshold in the attention weight map are analyzed and identified to determine the key micro-regions on the sealing surface that have the highest weight influencing sealing performance. The key micro-regions include the root of the sealing surface cutting edge, the junction of the transition arc and the sealing plane, and the inner and outer edges of the sealing strip. The location index information of the key micro-regions in the sealing surface-process association database is obtained. Based on the location index information of the key micro-region, the coordinate data of the morphological reference point at the corresponding location is extracted from the sealing surface-process association database, and the extracted coordinate data of multiple morphological reference points are combined to form a morphological reference point dataset.

5. The method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism according to claim 4, characterized in that, Based on the spatial coordinates of each point in the topographic reference point dataset, a convex hull polyhedron and a minimum volume covering ellipsoid are constructed respectively. The volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid is calculated and defined as the topographic envelope coefficient. The honing trajectory parameters are optimized using preset sealing performance indicators and morphological envelope coefficients to obtain optimized honing trajectory parameters, including: Obtain the topographic reference point dataset, and construct the convex hull polyhedron and the minimum volume covering ellipsoid that enclose the topographic reference point dataset based on the spatial coordinates of each point in the dataset. Calculate the volume of the convex hull polyhedron and the volume of the minimum volume covering ellipsoid. Find the ratio of the volume of the convex hull polyhedron to the volume of the minimum volume covering ellipsoid, and define the ratio as the morphology envelope coefficient. The morphological envelope coefficient and the preset sealing performance index are jointly constructed into a multi-objective optimization function. The honing trajectory parameters are iteratively optimized and solved with the minimization of the multi-objective optimization function value as the optimization objective. The honing trajectory parameters that satisfy the preset convergence condition of the multi-objective optimization function value are obtained as the optimized honing trajectory parameters.

6. The method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism according to claim 5, characterized in that, Based on the optimized honing trajectory parameters, the operating state of the sealing surface under various working conditions is simulated, and the helium leak detection, wear amount, and stress uniformity indicators are verified. If the verification results do not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weights, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, resulting in the final honing trajectory parameters, including: The optimized honing trajectory parameters are input into a pre-built multi-condition dynamic simulation model, which includes operating condition parameters under rated conditions and extreme conditions. The multi-condition dynamic simulation model is used to simulate the operating state of the sealing surface under various conditions, and the simulation values ​​of helium internal leakage, wear, and stress uniformity under the corresponding conditions are obtained. The simulated values ​​are compared with preset thresholds for helium internal leakage, wear, and stress uniformity to obtain the verification results for each index. If the verification results show that any simulation value does not meet the corresponding preset threshold, deviation information is generated based on the deviation between the unmet item and the threshold, and fed back to the trained convolutional neural network model to adjust its spatial attention weights, triggering the re-execution of steps 3 to 4 until all simulation values ​​meet the corresponding preset thresholds, and the honing trajectory parameters of the current iteration cycle are used as the final honing trajectory parameters.

7. The method for optimizing the honing trajectory of a semiconductor valve sealing surface based on a spatial attention mechanism according to claim 6, characterized in that, Also includes: Step 6: Based on the final honing trajectory parameters, generate the corresponding honing process scheme.

8. A semiconductor valve sealing surface honing trajectory optimization system based on spatial attention mechanism, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The database construction module is used to collect morphological feature data of the sealing surface of semiconductor valves, honing process parameters and corresponding sealing performance indicators, perform normalization preprocessing, and construct a sealing surface-process association database. The model building and training module is used to train a preset convolutional neural network model based on the sealing surface-process association database to obtain the trained convolutional neural network model. The key micro-area localization and dataset generation module is used to input the morphology data of the sealing surface to be processed into the trained convolutional neural network model, and locate the key micro-areas on the sealing surface through its spatial attention mechanism. Multiple morphological reference points corresponding to the key micro-regions are extracted from the sealing surface-process correlation database to form a morphological reference point dataset; The trajectory parameter optimization and solution module is used to construct a convex hull polyhedron and a minimum volume covering ellipsoid based on the spatial coordinates of each point in the topographic reference point dataset, calculate the volume ratio of the convex hull polyhedron to the minimum volume covering ellipsoid and define it as the topographic envelope coefficient. The honing trajectory parameters were optimized using preset sealing performance indicators and morphology envelope coefficients to obtain optimized honing trajectory parameters. The multi-condition simulation verification iteration module is used to simulate the operating state of the sealing surface under various conditions based on the optimized honing trajectory parameters, and to verify the helium detection internal leakage, wear amount and stress uniformity index. If the verification result does not reach the preset threshold, the generated deviation information is fed back to the trained convolutional neural network model to adjust its spatial attention weight, and steps 3 to 4 are re-executed until honing trajectory parameters that meet all preset thresholds are generated, and the final honing trajectory parameters are obtained. Based on the final honing trajectory parameters, the corresponding honing process scheme is generated.

Images