Semiconductor device flow channel polishing analysis method and system based on channel attention mechanism

Abstract

This invention provides a method and system for polishing semiconductor device flow channels based on a channel attention mechanism, relating to the field of precision semiconductor device processing technology. The method includes: constructing a flow channel-process correlation database; integrating geometric parameters, material parameters, extrusion honing process parameters, and performance data of the semiconductor device flow channel; performing normalization preprocessing to obtain standardized correlation data; constructing and training a convolutional neural network model incorporating a channel attention mechanism based on the standardized correlation data to obtain a trained convolutional neural network model; outputting the final extrusion honing process parameters for the target flow channel structure from the trained convolutional neural network model; and dynamically extracting flow channel topological anchor points at the ellipsoidal hollow valve cavity, conical variable diameter section, and arc-shaped turning corner of the target flow channel according to the final extrusion honing process parameters. This invention can achieve precise adaptation of flow channel polishing process parameters and stable performance compliance.

Description

Technical Field

[0001] This invention relates to the field of precision machining technology for semiconductor equipment, and in particular to a method and system for analyzing the polishing of semiconductor equipment flow channels based on the channel attention mechanism. Background Technology

[0002] In precision scenarios such as semiconductor chip manufacturing, wafer carrier processing, and fluid control pipelines for semiconductor etching equipment and thin film deposition equipment, the polishing quality of semiconductor equipment flow channels (such as fluid channels composed of ellipsoidal hollow valve cavities, conical variable diameter sections, and arc-shaped turning corners) determines the accuracy and stability of equipment operation. Among them, extrusion honing is currently the mainstream process for polishing semiconductor equipment flow channels, and the rationality of its process parameters directly affects key performance indicators such as the surface roughness of the flow channel, CV value error, and fluid residue.

[0003] For example, in the reaction gas channels of semiconductor etching equipment, insufficient channel polishing precision can lead to uneven gas flow, affecting etching uniformity. Conversely, poor polishing quality of the cooling channels in wafer carriers can cause uneven heat dissipation, affecting wafer processing yield. Current semiconductor equipment channel polishing analysis often suffers from the following technical defects: insufficient matching between extrusion honing process parameters and local channel morphology, making it difficult to achieve precise process parameter adaptation based on local channel morphological characteristics. This defect easily leads to over-polishing or under-polishing of channels, especially in complex local morphological areas such as ellipsoidal hollow valve cavities and arc-shaped corners, making it difficult to simultaneously meet the polishing precision requirements of different areas, affecting the smoothness of fluid flow. Furthermore, the lack of quantitative constraints on process parameters based on channel morphology results in significant fluctuations in channel performance indicators after polishing, making it difficult to stably reach preset standards and meet the high precision and high stability requirements of fluid channels in semiconductor chip manufacturing. Summary of the Invention

[0004] This invention provides a method and system for analyzing the polishing of semiconductor devices based on the channel attention mechanism, which can achieve accurate adaptation of the polishing process parameters and stable performance.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for analyzing the polishing of semiconductor device flow channels based on a channel attention mechanism, the method comprising: A flow channel-process correlation database is constructed, integrating the geometric parameters, material parameters, extrusion honing process parameters and performance data of semiconductor equipment flow channels, and performing normalization preprocessing to obtain standardized correlation data; Based on standardized associated data, a convolutional neural network model containing a channel attention mechanism is constructed and trained to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure. Based on the final extrusion honing process parameters, flow channel topology anchor points are dynamically extracted at the ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner of the target flow channel. A flow channel local morphology tensor field reflecting the local morphological distribution characteristics of the flow channel is constructed from the flow channel topology anchor points. Calculate the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel to obtain the flow channel morphological constraint tensor set used to constrain the polishing process; The flow channel morphology constraint tensor set and the final extrusion honing process parameters are input into a pre-trained multi-condition dynamic simulation model. Through simulation verification and iterative optimization, process parameters that reach the preset performance threshold and corresponding simulation verification results are formed. Based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, a final process scheme is formed, and polishing analysis results including a performance verification report are obtained.

[0006] Furthermore, a flow channel-process correlation database is constructed, integrating the geometric parameters, material parameters, extrusion honing process parameters, and performance data of semiconductor equipment flow channels. Normalization preprocessing is then performed to obtain standardized correlation data, including: Collect the geometric and material parameters of the flow channel of the target semiconductor device, construct a flow channel geometry-material data subset, and obtain the first data subset; Collect the extrusion honing process parameters and performance data corresponding to the flow channel geometry-material data subset, construct the process-performance data subset, and obtain the second data subset; The first data subset and the second data subset are associated and mapped according to the flow channel structure characteristics to form the original associated database; The data in the original relational database is preprocessed by normalization to eliminate the differences in dimensions and scales between different data sources, resulting in standardized relational data.

[0007] Furthermore, based on standardized correlated data, a convolutional neural network model incorporating a channel attention mechanism is constructed and trained to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure, including: The obtained standardized correlation data is used as the basic input data for model training to obtain the training dataset; An initial convolutional neural network model is constructed, and a channel attention mechanism is embedded in the initial convolutional neural network model to form a channel attention convolutional neural network model to be trained. The training dataset is input into the channel attention convolutional neural network model to be trained, and iterative training is performed with the flow channel polishing performance index as the optimization objective to obtain the trained convolutional neural network model. The geometric and material parameters of the target flow channel are input into the trained convolutional neural network model. The trained convolutional neural network model automatically predicts based on the key features weighted by the channel attention mechanism and outputs the final extrusion honing process parameters for the target flow channel structure.

[0008] Furthermore, based on the final extrusion honing process parameters, flow channel topological anchor points are dynamically extracted at the ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner of the target flow channel. A flow channel local morphology tensor field reflecting the local morphological distribution characteristics of the flow channel is constructed from these anchor points, including: The final extrusion honing process parameters for the target flow channel structure are obtained as the control benchmark for anchor point extraction, thus obtaining the process parameter constraint set. Based on the set of process parameter constraints, the first set of flow channel topological anchor points is dynamically extracted in the ellipsoidal hollow valve cavity region of the target flow channel according to the curvature change gradient, thus obtaining the valve cavity region topological anchor point set. Based on the process parameter constraint set, in the conical variable diameter section region of the target flow channel, the second set of flow channel topology anchor points is dynamically extracted according to the variable diameter ratio change curve to obtain the topology anchor point set of the variable diameter section region. Based on the set of process parameter constraints, the third set of flow channel topology anchor points is dynamically extracted in the circular arc turning corner region of the target flow channel according to the law of rotation radius variation, so as to obtain the topology anchor point set of the corner region. The topological anchor point set of the valve cavity region, the topological anchor point set of the variable diameter section region, and the topological anchor point set of the corner region are spatially fused to construct a local morphological tensor field of the flow channel that reflects the local morphological distribution characteristics of the flow channel, thus obtaining the local morphological tensor field of the flow channel.

[0009] Furthermore, the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel are calculated to obtain the flow channel morphological constraint tensor set used to constrain the polishing process, including: The constructed local morphological tensor field of the flow channel is obtained as the basic field data for feature quantity calculation, and the morphological tensor field to be analyzed is obtained. Tensor decomposition is performed on the morphological tensor field to be analyzed, and the principal eigenvalues ​​and principal eigendirections reflecting the local geometry of the flow channel are extracted to obtain the local morphological principal feature set. Based on the local morphological main feature set, the curvature tensor feature of the ellipsoidal hollow valve cavity region, the taper change rate feature of the conical variable diameter section region, and the rotation curvature feature of the circular arc turning corner region are calculated respectively to obtain the regionalized morphological feature set. By fusing the regionalized morphological feature sets, a comprehensive feature vector is constructed to characterize the overall morphological constraint capability of the flow channel, thus obtaining the comprehensive morphological constraint feature vector. By mapping the comprehensive morphological constraint feature vector to the polishing process parameter space, a flow channel morphological constraint tensor set that can quantitatively constrain the extrusion honing process parameters is formed, thus obtaining the flow channel morphological constraint tensor set used to constrain the polishing process.

[0010] Furthermore, the process parameters and corresponding simulation verification results that achieve the preset performance threshold through simulation verification and iterative optimization include: The flow channel morphology constraint tensor set and the final extrusion honing process parameters are used as joint input data for simulation verification and optimization to obtain the process parameter constraint combination to be verified. The process parameter constraints to be verified are input into the pre-trained multi-condition dynamic simulation model. The flow channel polishing performance is simulated under multiple preset conditions to obtain the initial simulation result set of the flow channel inner surface roughness, CV value error and fluid residue corresponding to each condition. Determine whether all performance indicators in the initial simulation result set have reached the preset performance threshold. If all have reached the threshold, output the current process parameters and simulation results as the compliance results to obtain the compliance process parameters and compliance simulation verification results. If there are performance indicators in the initial simulation results set that do not reach the preset performance threshold, then the final extrusion honing process parameters are adjusted in reverse according to the deviation between the non-compliant indicators and the preset performance threshold to form optimized process parameters, and thus obtain iteratively optimized process parameters. The iteratively optimized process parameters are re-inputted into the pre-trained multi-condition dynamic simulation model for iterative verification and adjustment until all performance indicators reach the preset performance threshold, thus obtaining the process parameters that have reached the preset performance threshold and the corresponding simulation verification results.

[0011] Furthermore, based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, a final process scheme is formed, resulting in polishing analysis results including a performance verification report, including: The process parameters that reach the preset performance threshold and the corresponding simulation verification results are used as the core components of the final process scheme to obtain the combination of process parameters and verification data to be output. The process parameters to be output are combined with the process parameters in the verification data and encapsulated to form a standard process specification containing the extrusion honing process parameters and their control range, thus obtaining a standard process specification file. The simulation verification results in the combination of the process parameters to be output and the verification data are classified and organized according to the performance indicators of the inner surface roughness of the flow channel, CV value error and fluid residue, and a standard verification data table for each performance indicator is generated to obtain a verification data summary table. By linking and integrating the standard process specification documents with the verification data summary table, a comprehensive performance verification report is generated, which includes process parameters, verification results, and performance compliance conclusions. The polishing analysis results containing the performance verification report are then obtained.

[0012] Secondly, a semiconductor device flow channel polishing analysis system based on the channel attention mechanism includes: The database construction module is used to build a flow channel-process correlation database, which integrates the geometric parameters, material parameters, extrusion honing process parameters and performance data of semiconductor equipment flow channels, performs normalization preprocessing, and obtains standardized correlation data. The model training and parameter prediction module is used to construct and train a convolutional neural network model containing a channel attention mechanism based on standardized correlation data, so as to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure. The morphological tensor field construction module is used to dynamically extract the topological anchor points of the target flow channel at the ellipsoidal hollow valve cavity, the conical variable diameter section and the arc turning corner, respectively, based on the final extrusion honing process parameters. The flow channel topological anchor points are used to construct a flow channel local morphological tensor field that reflects the local morphological distribution characteristics of the flow channel. The morphological constraint tensor calculation module is used to calculate the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel, and obtain the flow channel morphological constraint tensor set used to constrain the polishing process. The simulation verification and iterative optimization module is used to input the flow channel morphology constraint tensor set and the final extrusion honing process parameters into the pre-trained multi-condition dynamic simulation model. Through simulation verification and iterative optimization, process parameters that reach the preset performance threshold and corresponding simulation verification results are formed. The polishing analysis result generation module generates a final process scheme based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, and obtains polishing analysis results including a performance verification report.

[0013] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: Because a convolutional neural network model incorporating channel attention is used to train and predict standardized correlated data, the technical problems of insufficient matching between extrusion honing process parameters and local channel morphology, and the difficulty in achieving precise adaptation based on experience, are overcome. This allows for the automatic output of optimal extrusion honing process parameters for the target channel structure, improving the matching accuracy between process parameters and channel geometry. Furthermore, by dynamically extracting channel topological anchor points and constructing a local channel morphology tensor field at the ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner based on the final extrusion honing process parameters, the technical problems of insufficient matching between extrusion honing process parameters and local channel geometry are overcome. The technical problem of over-polishing or under-polishing is difficult to balance the precision requirements of different parts when polishing complex areas. This allows for a refined characterization of the local morphology of the flow channel, providing a precise spatial morphological constraint basis for the polishing process. By inputting the flow channel morphology constraint tensor set and the final extrusion honing process parameters into the multi-condition dynamic simulation model and performing iterative optimization, the technical problem of large fluctuations in the performance indicators of the flow channel after polishing and difficulty in stably reaching the preset standards is overcome. This results in process parameters and simulation verification results that reach the preset performance threshold, improving the consistency and stability of the flow channel polishing quality. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of a semiconductor device flow channel polishing analysis method based on channel attention mechanism provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a semiconductor device flow channel polishing analysis system based on the channel attention mechanism provided in an embodiment of the present invention. Detailed Implementation

[0018] 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.

[0019] like Figure 1 As shown, embodiments of the present invention propose a semiconductor device flow channel polishing analysis method based on a channel attention mechanism, the method comprising the following steps: Step 1: Construct a flow channel-process correlation database, integrate the geometric parameters, material parameters, extrusion honing process parameters and performance data of semiconductor equipment flow channels, perform normalization preprocessing, and obtain standardized correlation data; Step 2: Based on standardized associated data, construct and train a convolutional neural network model that includes a channel attention mechanism to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure. Step 3: Based on the final extrusion honing process parameters, dynamically extract the topological anchor points of the flow channel at the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner of the target flow channel. Construct a flow channel local morphology tensor field that reflects the local morphological distribution characteristics of the flow channel from the flow channel topological anchor points. Step 4: Calculate the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel to obtain the flow channel morphological constraint tensor set used to constrain the polishing process. Step 5: Input the flow channel morphology constraint tensor set and the final extrusion honing process parameters into the pre-trained multi-condition dynamic simulation model. After simulation verification and iterative optimization, process parameters that reach the preset performance threshold and corresponding simulation verification results are formed. Step 6: Based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, form the final process scheme and obtain the polishing analysis results including the performance verification report.

[0020] In this embodiment of the invention, by constructing a channel-process association database and using a convolutional neural network model with a channel attention mechanism for training and prediction, precise matching between extrusion honing process parameters and channel geometric features is achieved, effectively replacing the traditional experience-based manual parameter optimization method. By dynamically extracting channel topological anchor points at ellipsoidal hollow valve cavities, conical variable diameter sections, and arc-shaped turning corners and constructing a local channel morphology tensor field, a refined characterization of the complex channel morphology is achieved, providing a spatial morphological constraint basis for the polishing process. By calculating the characteristic quantities of the morphology tensor field to form a channel morphology constraint tensor set, and combining iterative optimization with a multi-condition dynamic simulation model, it is ensured that process parameters can reach preset performance thresholds under various conditions. Finally, a polishing analysis result including a performance verification report is generated, effectively improving the stability and consistency of channel polishing quality, shortening the process optimization cycle, reducing R&D costs, and meeting the stringent requirements of semiconductor equipment for high precision and high cleanliness.

[0021] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Collect the geometric and material parameters of the flow channel of the target semiconductor device, construct a flow channel geometry-material data subset, and obtain the first data subset. Specifically, this includes: performing a full-surface scan of the flow channel of the target semiconductor device to obtain three-dimensional point cloud data of the ellipsoidal hollow valve cavity region; extracting multiple transverse sections from the point cloud data along different azimuth angles; extracting the radius of curvature values ​​of each section; fitting the radius of curvature of each section according to the section position to generate curvature distribution data; for the conical variable diameter section region, setting multiple sampling positions at equal steps along the flow channel axis; measuring the cross-sectional diameter values ​​at each sampling position; calculating the diameter change between adjacent sampling positions and dividing it by the sampling interval to obtain the variable diameter ratio; and arranging the variable diameter ratios of each sampling position in order of axial position. The curve of the variable diameter ratio is generated; for the circular arc turning corner region, multiple arc segments are extracted along the direction of the turning arc, and the corresponding gyration radius value and central angle span of each arc segment are measured, and the distribution range of the central angle is recorded; at the same time, the hardness value, elastic modulus value and surface energy characteristic data of the flow channel matrix material are obtained; the above-collected curvature distribution data, variable diameter ratio curve, gyration radius value and central angle distribution range are classified and organized into three regions: ellipsoidal hollow valve cavity, conical variable diameter segment and circular arc turning corner, and stored in the data fields of the corresponding regions respectively. The hardness value, elastic modulus value and surface energy characteristic data are used as material parameters and associated with each region to construct a flow channel geometry-material data subset, and the first data subset is obtained.

[0022] Step 1.2: Collect the extrusion honing process parameters and performance data corresponding to the flow channel geometry-material data subset, construct the process-performance data subset, and obtain the second data subset. Specifically, this includes: for each batch of extrusion honing processing corresponding to the flow channel geometry-material data set, extract the abrasive concentration percentage, feed rate value, polishing pressure value, and single processing time of that batch from the process record system of the extrusion honing equipment to form a process parameter record; at the same time, obtain the measured average value of the inner surface roughness of the flow channel after processing, the measured deviation percentage of the CV value error, and the measured mass percentage of the fluid residue from the flow channel performance test report to form a performance data record; associate the process parameter record and the performance data record by processing batch, and label the corresponding measured performance data of each region with three regions as the granularity: ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner, to construct a process-performance data subset containing multiple batches and multiple region correspondences, and obtain the second data subset.

[0023] Step 1.3 involves mapping the first and second data subsets according to their flow channel structure characteristics to form an original association database. Specifically, this includes mapping the first and second data subsets together. First, the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner are used as primary association keys. These primary keys determine the physical structure region to which the data belongs, ensuring that all geometric parameters, material parameters, process parameters, and performance data within the same flow channel structure region are grouped into the same group, establishing a structural alignment benchmark for subsequent matching. Then, within the same structural region, the unique identifier of the processing batch is used as a secondary association key. This secondary association key is used to identify each data record within the same batch. The time-to-production correspondence ensures that the geometric material data corresponding to a batch can be accurately linked to the process performance data of the same batch. During the matching process, a subset of data in the same structural region is first filtered out using a primary association key. Then, within this subset, a search and matching process is performed based on a secondary association key. The geometric parameters and material parameters of the region in the first data subset are verified one-to-one with the process parameters and performance data of the same batch and region in the second data subset. After successful verification, the successfully matched data records are merged into a complete data entry. This data entry contains information in six dimensions: region identifier, batch identifier, geometric parameters, material parameters, process parameters, and performance data, forming an original association database covering multiple structural regions and multiple process batches.

[0024] Step 1.4 involves performing normalization preprocessing on the data in the original relational database to eliminate dimensional and scale differences between different data sources, resulting in standardized relational data. Specifically, this includes: performing normalization preprocessing on the data in the original relational database; extracting the original numerical ranges of four data items: geometric parameters, material parameters, process parameters, and performance data; identifying the maximum and minimum values ​​of each data item; and using the max-min normalization method, calculating the difference between the minimum and maximum values ​​for each value in each data item category, dividing by the difference between the maximum and minimum values, and mapping the calculation result to [0, 1, 2, 3]. , The numerical range of 1]; during the normalization process, the data grouping structure of the three regions of ellipsoidal hollow valve cavity, conical variable diameter section and circular arc turning corner remains unchanged, while the original correlation between each data item is preserved, and the data distribution deviation caused by different dimensions and scale differences between geometric parameters, material parameters, process parameters and performance data is eliminated to obtain standardized correlation data.

[0025] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Obtain the standardized correlation data as the basic input data for model training to obtain the training dataset. Specifically, this includes: obtaining the standardized correlation data, using the geometric and material parameters of the three structural regions—ellipsoidal hollow valve cavity, conical variable diameter section, and arc-shaped turning corner—as input features, using the extrusion honing process parameters of the corresponding regions as output labels, and using the surface roughness of the flow channel, CV value error, and fluid residue as optimization objective constraints; dividing the data into a training set and a validation set according to a preset ratio, where the preset ratio is to randomly divide all data samples in an 8:2 ratio, that is, 80% of the samples are used as the training set for learning and fitting model parameters, and 20% of the samples are used as the validation set for performance evaluation and hyperparameter adjustment during model training. When dividing, ensure that both the training set and the validation set cover samples from the three structural regions—ellipsoidal hollow valve cavity, conical variable diameter section, and arc-shaped turning corner—and that the distribution ratio of samples from each region in the training set and the validation set is consistent with the original data, thus constructing a training dataset for model training.

[0026] Step 2.2: Construct an initial convolutional neural network model. Embed a channel attention mechanism in the initial convolutional neural network model to form a channel attention convolutional neural network model to be trained. Specifically, this includes: constructing an initial convolutional neural network model, which is an improved version of the residual network architecture to adapt to the feature extraction requirements of multidimensional heterogeneous data in the semiconductor equipment flow channel polishing analysis scenario; first, initialize the network model, set the number of input layer nodes to match the total dimension of the normalized geometric parameters and material parameters of the three structural regions: ellipsoidal hollow valve cavity, conical variable diameter section, and arc turning corner; and set the number of output layer nodes to match the output dimensions of the twelve process parameters corresponding to the three structural regions: abrasive concentration, feed rate, polishing pressure, and processing time.

[0027] This model structure consists of an input layer, multiple residual convolutional blocks, a channel attention mechanism module, a global pooling layer, and a fully connected output layer cascaded together. The input layer receives data and first reshapes the features, arranging the parameters of each region according to their region dimensions to form a three-dimensional feature tensor. Each residual convolutional block consists of a convolutional layer, a batch normalization layer, an activation function layer, and skip connections. The first residual convolutional block performs shallow feature extraction on the input feature tensor, capturing the independent geometric features of each region. Subsequent residual convolutional blocks gradually expand the receptive field, extracting deeper features that are interconnected between regions. In each residual convolutional block, skip connections directly add the input features to the convolutional output features before feeding them into the next convolutional block. The module alleviates the gradient vanishing problem in deep network training. The channel attention mechanism module is embedded after the last residual convolutional block. It performs global average pooling and global max pooling on the feature map output by the residual convolutional block to generate two channel-dimensional description vectors. The two description vectors are then input into a shared multilayer perceptron for nonlinear transformation. The two transformed feature vectors are added element by element and then activated by an activation function to generate weight coefficients for each feature channel. The weight coefficients are multiplied with the original feature map channel by channel to automatically enhance geometric feature channels that are highly correlated with the polishing quality of the flow channel and suppress redundant feature channels, thereby realizing the selection of process parameters based on the local morphological features of the flow channel.

[0028] The global pooling layer compresses the attention-weighted feature map into a one-dimensional feature vector. The fully connected output layer maps this feature vector to the abrasive concentration, feed rate, polishing pressure, and processing time corresponding to the three structural regions. At the same time, a dropout layer is introduced before the fully connected output layer to prevent overfitting. The architecture described above can retain the multi-scale features of the flow channel geometry while achieving accurate selection of key process parameters through the attention mechanism. Compared with traditional convolutional neural networks, it effectively improves the sensitivity to local morphological features such as abrupt curvature changes in ellipsoidal valve cavities, taper changes in conical variable diameter sections, and differences in the radius of rotation of arc corners, thereby enhancing the model's generalization ability for different flow channel structures and the accuracy of process parameter prediction.

[0029] Step 2.3 involves inputting the training dataset into the channel attention convolutional neural network model to be trained, and iteratively training it with the flow channel polishing performance index as the optimization objective to obtain the trained convolutional neural network model. Specifically, this includes: inputting the obtained training dataset into the channel attention convolutional neural network model constructed in Step 2.2; using forward propagation to input the geometric and material parameters of the three structural regions in the training dataset—the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner—layer by layer through feature extraction from residual convolutional blocks, feature weighting from the channel attention mechanism module, and feature compression from the global pooling layer; finally, the model outputs the predicted abrasive concentration, feed rate, polishing pressure, and processing time through a fully connected output layer; calculating the mean square error between the predicted values ​​and the corresponding actual process parameters in the training dataset as the main loss function; and simultaneously adding the deviations between the predicted and actual values ​​of the flow channel inner surface roughness, CV value error, and fluid residue as auxiliary constraints to the loss function, forming a comprehensive loss function, the expression of which is: Among them, the main loss function The formula used to measure the accuracy of process parameter predictions is as follows:

[0030] In the formula, This indicates the number of structural regions, with a value of 3, corresponding to the three regions: the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner. This indicates the number of process parameters corresponding to each region, with a value of 4, corresponding to abrasive concentration, feed rate, polishing pressure, and processing time. The model represents the first i The first region j Predicted values ​​of each process parameter; Indicates the first in the training dataset i The first region j Actual values ​​of each process parameter; auxiliary constraint items The formula used to constrain the consistency between the performance indicators of the polished flow channel and the target value is as follows:

[0031] in, This indicates the number of performance indicators, with a value of 3, corresponding to the surface roughness of the flow channel, CV value error, and fluid residue. The model represents the first i The first region k Predicted values ​​for each performance indicator; Indicates the first i The first region k Target values ​​for each performance metric; These are the weighting coefficients for the auxiliary constraint terms, used to balance the contributions of the main loss function and the auxiliary constraint terms to the overall loss function. The value of is typically between 0.01 and 0.1, and the specific value can be determined by grid search or Bayesian optimization based on the performance of the comprehensive loss function on the validation set. The backpropagation algorithm is used to calculate the gradient of the comprehensive loss function with respect to the weights of each layer from the output layer to the input layer. An adaptive moment estimation optimizer is used to iteratively update the network weights based on the calculated gradients. During each update, a batch of data is randomly selected for training, with a batch size of 32 samples. After training all batches for each training cycle, the validation set is used to evaluate the performance of the current model. The comprehensive loss function value on the validation set is calculated, and its trend is recorded. When the comprehensive loss function value on the validation set no longer decreases or shows an upward trend for five consecutive training cycles, an early stopping mechanism is triggered to terminate training. The network weights saved at this point are used as the final training result, resulting in the completed convolutional neural network model.

[0032] Step 2.4: Input the geometric and material parameters of the target flow channel into the trained convolutional neural network model. The trained convolutional neural network model automatically predicts based on the key features weighted by the channel attention mechanism, and outputs the final extrusion honing process parameters for the target flow channel structure. Specifically, this includes: acquiring the curvature distribution data of the ellipsoidal hollow valve cavity of the target flow channel, the diameter ratio change curve of the conical variable diameter section, the gyration radius of the circular arc turning corner, and the hardness, elastic modulus, and internal surface energy characteristics of the flow channel matrix material. The above data are then processed according to the normalization preprocessing method used in Step 1.4. The mapping parameters are subjected to the same normalization transformation to obtain normalized input data. This input data is then grouped into three regions: ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner. Each group contains the geometric and material parameters of that region. The data of the three regions are arranged in the order of the regions to form a three-dimensional input feature tensor. The first dimension of this three-dimensional input feature tensor represents the region category, the second dimension represents the feature parameter type within that region, and the third dimension represents the specific numerical sequence of the feature parameters. This three-dimensional structure completely preserves the spatial correlation between the geometric shape and material properties of each region.

[0033] This 3D input feature tensor is fed into the trained convolutional neural network model. The model extracts features from the input feature tensor layer by layer through residual convolutional blocks. The first residual convolutional block extracts the independent shallow geometric features of each region, while subsequent residual convolutional blocks extract the deep features that are interconnected between regions within an increasingly larger receptive field. The channel attention mechanism module processes the feature map output by the last residual convolutional block. F We perform channel-dimensional weight recalibration to learn the attention weight coefficients for each feature channel. Furthermore, the key features are enhanced and redundant features are suppressed through a channel weighting formula: In the formula C This represents the total number of channels in the feature map. For the first Attention weight coefficients for each feature channel, For the first Feature map of each channel For channel-by-channel multiplication, The result is a weighted fused feature map. After weighting, the model automatically enhances the geometric and material feature channels that are most correlated with polishing quality, while suppressing the interference of redundant feature channels on the prediction results. The global pooling layer then uses the weighted fused feature map... The feature vector is compressed into a one-dimensional feature vector, which integrates the key morphological feature information of the three regions. The fully connected output layer maps this one-dimensional feature vector to the abrasive concentration, feed rate, polishing pressure and processing time corresponding to the three structural regions, for a total of twelve process parameter values, forming the final extrusion honing process parameters for the target flow channel structure.

[0034] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Obtain the final extrusion honing process parameters for the target flow channel structure as the control benchmark for anchor point extraction, thus obtaining a process parameter constraint set. Specifically, this includes obtaining the final extrusion honing process parameters for the target flow channel structure. These parameters include the abrasive concentration, feed rate, polishing pressure, and processing time corresponding to the three regions: the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner. These parameters are used as the control benchmark for anchor point extraction to determine the density, distribution strategy, and screening threshold for anchor point extraction in each region, forming a process parameter constraint set. This constraint set is denoted as... ,in These respectively represent the ellipsoidal hollow valve cavity region, the conical variable diameter section region, and the circular arc turning corner region. These represent abrasive concentration, feed rate, polishing pressure, and processing time, respectively.

[0035] Step 3.2: Based on the process parameter constraint set, dynamically extract the first set of flow channel topological anchor points according to the curvature change gradient within the ellipsoidal hollow valve cavity region of the target flow channel to obtain the valve cavity region topological anchor point set. Specifically, this includes: constructing a three-dimensional spatial coordinate system within the ellipsoidal hollow valve cavity region of the target flow channel based on the obtained process parameter constraint set; performing grid discretization processing on the inner surface of the flow channel in this region along both the circumferential and axial directions, dividing it into M equal intervals along the circumferential direction and N equal intervals along the axial direction, forming M×N grid units, with each grid node corresponding to a spatial coordinate point; for each grid node, calculating the Gaussian curvature and average curvature of the surface at that node, and further calculating the curvature change gradient magnitude to characterize the degree of curvature change of the valve cavity surface along different directions at that node, generating a curvature change gradient map covering the entire valve cavity region; where nodes Curvature change gradient magnitude at The calculation formula is: In the formula, For nodes Gaussian curvature value at that point, u and v These are the two parameterized directions of the valve cavity surface, corresponding to the circumferential and axial curve coordinates, respectively. and Gaussian curvature at u direction and v Partial derivatives in direction.

[0036] Based on the process parameter constraint set, a curvature gradient threshold is set for the abrasive concentration and polishing pressure parameters in the valve cavity region. When the curvature change gradient magnitude calculated at a certain node When the value is greater than or equal to the threshold, the node is identified as a location of significant morphological change, and a flow channel topology anchor point is extracted at this location. Simultaneously, to ensure uniform anchor point distribution and complete coverage of the valve cavity region, anchor points are added circumferentially and axially at equal arc length intervals in areas where curvature changes gently and the threshold condition is not triggered. All extracted anchor points, their spatial coordinates, region identifiers, and corresponding Gaussian curvature values ​​are recorded as a valve cavity region topology anchor point set, denoted as […]. ,in This refers to the index number of the anchor point within the valve cavity area.

[0037] Step 3.3: Based on the process parameter constraint set, dynamically extract the second set of flow channel topology anchor points within the conical variable diameter section region of the target flow channel according to the variable diameter ratio change curve, obtaining the topology anchor point set for the variable diameter section region. Specifically, this includes: based on the obtained process parameter constraint set, establishing a one-dimensional parametric coordinate system along the flow channel axis direction within the conical variable diameter section region of the target flow channel, assuming the axis direction parameter is... s The value range starts from the beginning of the variable diameter section. To the end The diameter at each cross-section of the variable diameter section is expressed as an axial position parameter. s The function, denoted as D ( s This function is obtained by fitting the diameter of the cross-section sampled at equal intervals along the axis of the three-dimensional model of the flow channel; for the diameter function D ( s Find the first derivative to obtain the curve of the change in diameter ratio. The curve reflects the rate at which the cross-sectional diameter changes with the axis at different locations of the variable diameter section.

[0038] Based on the process parameter constraint set, the diameter ratio sampling interval Δ is set for the feed rate and processing time parameters of the diameter-changing section. s Sampling points are set at equal intervals along the axis starting from the beginning, and the results are calculated for each sampling point sequentially. Variable diameter at point ,in The total number of sampling points is Δ, and the distance between adjacent sampling points is the set sampling interval. s ; Calculate the change in diameter ratio between adjacent sampling points When the change exceeds a preset threshold When this occurs, it indicates a significant change in the diameter ratio within that interval. At this time, respectively at... and Topological anchor points for the flow channel are extracted at two sampling points. To ensure uniform distribution and complete coverage of the variable diameter section, anchor points are added at equal intervals in areas where the diameter change rate is gradual and the threshold condition is not triggered. All extracted anchor points, their axial coordinates, cross-sectional diameter information, and diameter change rate values ​​are recorded to form a topological anchor point set for the variable diameter section, denoted as [set name missing]. ,in β This refers to the index number of the anchor points within the variable diameter section area.

[0039] Step 3.4: Based on the process parameter constraint set, dynamically extract the third set of flow channel topology anchor points within the circular arc turning corner region of the target flow channel according to the variation law of the turning radius, obtaining the topology anchor point set of the corner region. Specifically, this includes: based on the obtained process parameter constraint set, establishing a parametric model of the rotating surface within the circular arc turning corner region of the target flow channel, assuming the turning direction parameter is... θ The value range starts from the corner. To the termination angle Let the busbar direction parameter be... η The value range starts from the inside of the corner. To the outside The corner surface is unfolded into a two-dimensional parameter domain along the rotation direction, where the position of each parameter is... For a spatial point on the corner surface, calculate the radius of gyration at each parameter location. This value represents the perpendicular distance from the point to the axis of rotation, and further calculates the differential rate of change of the radius of rotation along the direction of rotation. It is used to characterize the curvature variation of a corner surface at different rotation angles.

[0040] Based on the process parameter constraint set, a threshold for the rate of change of the radius of rotation is set for the abrasive concentration and polishing pressure parameters in the corner region. The absolute value of the differential rate of change of the radius of gyration along the direction of gyration at a certain parameter position. When the value is greater than or equal to the threshold, the location is identified as a significant morphological change location, and a flow channel topology anchor point is extracted at this location. Simultaneously, to ensure uniform anchor point distribution and complete coverage of the corner area, anchor points are added along the rotation direction at equal arc length intervals in areas where the radius of rotation changes gently and the threshold condition is not triggered. All extracted anchor points, along with their parameter coordinates, spatial coordinates, radius of rotation values, and rate of change of radius of rotation information, are recorded to form a corner area topology anchor point set, denoted as […]. ,in γ This refers to the index number of the anchor point within the corner area.

[0041] Step 3.5 involves spatially fusing the topological anchor point sets of the valve cavity region, the variable diameter section region, and the corner region to construct a local morphological tensor field reflecting the local morphological distribution characteristics of the flow channel. Specifically, this includes: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The obtained topological anchor point set of the variable diameter segment region With the obtained topological anchor point set of the corner region To perform spatial fusion, all anchor points in the three anchor point sets are first mapped to the same spatial coordinate system to ensure that the spatial coordinates of each anchor point have a unified reference datum. Let the spatial coordinates of any anchor point 'a' after fusion be... Each anchor point is assigned an identifier for its region and corresponding curvature, diameter ratio, or radius of gyration attribute values; then, the merged complete anchor point set is used... As sampling points, a continuous tensor field is constructed in the three-dimensional space of the flow channel. ,in Let x be the position coordinates of any point within the flow channel space; using the radial basis function interpolation method, the morphological feature information at the discrete anchor points is made continuous, and the formula for calculating the second-order tensor value T(x) at any position x in the tensor field is:

[0042] In the formula For fusion anchor set General continuous index of anchor points, For the first fusion anchor point set The spatial coordinates of each anchor point The radial basis functions are in Gaussian form. ,in This represents the Euclidean distance between the spatial location point and the anchor point. These are shape parameters used to control the influence range of the radial basis functions; For the first A second-order tensor weight matrix is ​​constructed at each anchor point based on its curvature, rate of change of diameter, or radius of gyration attribute values. The diagonal elements of this weight matrix reflect the principal curvature or principal rate of change of diameter at the anchor point along different directions, while the off-diagonal elements reflect the coupling relationship between different directions. Through the above interpolation calculation, a local morphological tensor field of the flow channel continuously distributed in the three-dimensional space of the flow channel is obtained. This tensor field can fully reflect the local morphological distribution characteristics of the three regions—the ellipsoidal hollow valve cavity, the conical variable diameter region, and the circular arc turning corner—at different spatial locations.

[0043] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Obtain the constructed local morphological tensor field of the flow channel as the basic field data for feature quantity calculation, and obtain the morphological tensor field to be analyzed. Specifically, this includes: obtaining the constructed local morphological tensor field of the flow channel. This tensor field is a vector field continuously distributed within the three-dimensional space of the flow channel, with each spatial location point... Each location corresponds to a second-order tensor value. This second-order tensor fully describes the curvature shape and spatial orientation characteristics of the location along different spatial directions through its nine components, including the magnitude and direction of the principal curvature, as well as the degree of curvature and anisotropy of the surface in different directions. This tensor field is used as the basic field data for subsequent feature quantity calculations to ensure that the extraction of all morphological feature quantities is based on this tensor field, thus obtaining the morphological tensor field to be analyzed.

[0044] Step 4.2: Perform tensor decomposition on the morphological tensor field to be analyzed, extract the principal eigenvalues ​​and principal eigendirections reflecting the local geometry of the flow channel, and obtain the local morphological principal feature set. Specifically, this includes: performing tensor decomposition on the morphological tensor field to be analyzed, for each discrete sampling point in the three-dimensional space of the flow channel. The second-order tensor at the location The eigenvalue decomposition method is used to decompose it into three eigenvalues ​​and three corresponding eigendirections. The specific decomposition process involves solving the characteristic equation. ,in λ For eigenvalues, For the corresponding eigenvectors, we obtain three eigenvalues. With the corresponding feature vector ; where the largest eigenvalue Reflects the maximum curvature intensity along the principal direction at this location, the second largest eigenvalue. The minimum eigenvalue reflects the curvature shape intensity along the secondary principal direction. Reflecting the curvature intensity along a third direction; after decomposing the eigenvalues ​​at each location point, sorting them from largest to smallest, and selecting the first two principal eigenvalues. , With the corresponding main feature direction , This process generates principal feature values ​​and principal feature directions that reflect the local geometry of the flow channel. The principal feature values ​​and principal feature directions of all sampled locations are then aggregated according to their spatial location indices to construct a local morphological principal feature set containing the spatial coordinates, principal feature values, and principal feature directions of each location.

[0045] Step 4.3: Based on the local morphological principal feature set, calculate the curvature tensor feature of the ellipsoidal hollow valve cavity region, the taper change rate feature of the conical variable diameter section region, and the rotation curvature feature of the circular arc rotation corner region, respectively, to obtain a regionalized morphological feature set. Specifically, this includes: based on the obtained local morphological principal feature set, calculating the corresponding morphological feature quantities for the ellipsoidal hollow valve cavity region, the conical variable diameter section region, and the circular arc rotation corner region, respectively. For the ellipsoidal hollow valve cavity region, calculate the curvature tensor feature quantity. Let the principal feature value at any point within the valve cavity region be... and ,in p Given the index of a location point within the valve cavity region, the curvature tensor characteristic of that point is... The curvature tensor features of all points within the valve cavity region are statistically analyzed to obtain the distribution set of curvature tensor features of the valve cavity region. For the conical variable diameter section region, the taper rate of change feature is calculated, assuming the positional parameter along the axial direction within the variable diameter section region is... s The cross-sectional diameter function at this location is D ( s Step 4.2 yields the angle between the secondary principal feature direction and the axis direction at each location point in the region. ,in q Given the index of a location point within the variable diameter section region, the characteristic quantity of the rate of taper change at that point is... .

[0046] in, This represents the diameter ratio at that location. Using the cosine of the angle between the secondary principal characteristic direction and the axial direction, the taper change rate characteristic of all points within the variable diameter section is statistically analyzed to obtain the distribution set of the taper change rate characteristic of the variable diameter section. For the circular arc turning corner region, the curvature characteristic is calculated, assuming the radius of gyration at any point within the corner region is... Step 4.2 yields the angle between the principal feature direction and the rotation direction at that location. ,in r Given the index of a location point within the corner region, the rotation curvature characteristic of that point is... In the formula, Let be the differential rate of change of the radius of gyration at that location along the direction of gyration. Using the sine value of the angle between the main feature direction and the rotation direction, the rotation curvature feature quantities of all positions in the corner area are statistically analyzed to obtain the distribution set of rotation curvature feature quantities in the corner area; the distribution sets of the curvature tensor feature quantities of the valve cavity area, the distribution sets of the taper change rate feature quantities of the variable diameter section area, and the distribution sets of rotation curvature feature quantities of the corner area are merged to obtain the regionalized morphological feature quantity set.

[0047] Step 4.4 involves fusing the regionalized morphological feature sets to construct a comprehensive feature vector characterizing the overall morphological constraint capability of the flow channel. This comprehensive morphological constraint feature vector is obtained by fusing the obtained regionalized morphological feature sets, first extracting the statistical mean of the curvature tensor feature quantities of the ellipsoidal hollow valve cavity region. with standard deviation Statistical mean of the characteristic quantity of the taper change rate in the conical variable diameter section region with standard deviation Statistical mean of the curvature characteristic of the arc-shaped corner region with standard deviation The six statistics mentioned above are combined in regional order to form a one-dimensional feature vector. Simultaneously, the correlation between the feature quantities in the three regions is incorporated into the fusion process in the form of a covariance matrix. This constructs a comprehensive feature vector characterizing the overall morphological constraint capability of the flow channel, denoted as... ,in Given the covariance matrix among the three regional feature quantities, we obtain the comprehensive morphological constraint feature vector.

[0048] Step 4.5: Map the comprehensive morphological constraint feature vector to the polishing process parameter space to form a flow channel morphological constraint tensor set that can quantitatively constrain the extrusion honing process parameters, thus obtaining the flow channel morphological constraint tensor set used to constrain the polishing process. Specifically, this includes: mapping the obtained comprehensive morphological constraint feature vector... Mapping to the polishing process parameter space, a pre-trained regression mapping network model is used for mapping. This regression mapping network model is a multi-layer fully connected neural network structure, including an input layer, two hidden layers, and an output layer. The number of nodes in the input layer is set to the dimension of the comprehensive morphological constraint feature vector, and the number of nodes in the output layer is set to 12, corresponding to the four process parameters of abrasive concentration, feed rate, polishing pressure, and processing time for the three regions: the ellipsoidal hollow valve cavity, the conical variable diameter section, and the arc turning corner.

[0049] The comprehensive morphological constraint feature vector The input features are fed into the regression mapping network model. The input layer receives the feature vector and passes it to the first hidden layer. The first hidden layer performs a linear transformation using the weight matrix and bias term, followed by a nonlinear mapping using an activation function. The second hidden layer further extracts higher-order features. Finally, the output layer outputs the process parameter correction coefficient matrix for each of the three regions. This matrix is ​​a 3×4 tensor, denoted as Δ, where the row index corresponds to the three regions: the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner; and the column index corresponds to the four process parameters: abrasive concentration, feed rate, polishing pressure, and processing time. The final extrusion honing process parameters obtained in step 2.4 are denoted as... It is also in 3×4 tensor form, and will Multiply by Δ element by element, i.e. Where ⊙ represents the Hadamard product operation, the final process parameter tensor obtained after morphological constraint correction is obtained. This tensor is the set of flow channel morphology constraint tensors that can quantitatively constrain the extrusion honing process parameters, denoted as... The tensor set fully includes the corrected abrasive concentration, feed rate, polishing pressure, and processing time for each of the three regions, and is used for subsequent simulation verification and iterative optimization.

[0050] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Using the channel morphology constraint tensor set and the final extrusion honing process parameters as joint input data for simulation verification and optimization, the process parameter constraint combination to be verified is obtained. Specifically, this includes: acquiring the obtained channel morphology constraint tensor set and the output final extrusion honing process parameters. The channel morphology constraint tensor set is the process parameter tensor after morphology constraint correction, with a data structure of a 3×4 matrix. The row indices correspond to three regions: the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner. The column indices correspond to four process parameters: abrasive concentration, feed rate, polishing pressure, and processing time. The final extrusion honing process parameters are the basic process parameter tensors without morphology constraint correction. The data structure is also a 3×4 matrix, with the row and column indices having the same meaning as the flow channel morphology constraint tensor set. The two tensors are combined according to the region dimension. Specifically, for each region, the modified process parameters in the flow channel morphology constraint tensor set corresponding to that region are paired with the basic process parameters in the final extrusion honing process parameters to form a data entry containing a binary pair of basic parameters and modified parameters. The binary pairs of the three regions are aggregated in the order of the regions to construct a joint input data structure. This structure is presented in the form of a three-dimensional tensor with a dimension of 3×4×2. The third dimension stores the basic process parameters and the modified process parameters respectively, resulting in the process parameter constraint combination to be verified.

[0051] Step 5.2: Input the process parameter constraints to be verified into the pre-trained multi-condition dynamic simulation model, and perform flow channel polishing performance simulation under multiple preset conditions to obtain the initial simulation result set of flow channel inner surface roughness, CV value error and fluid residue for each condition. Specifically, this includes: constructing the multi-condition dynamic simulation model to be trained. This model is built based on a physical information neural network architecture, which deeply integrates fluid dynamics control equations with data-driven methods to adapt to the simulation requirements of multi-physics coupling in the flow channel polishing process of semiconductor equipment; collecting historical flow channel polishing case data, including the measured values ​​of flow channel inner surface roughness, CV value error and fluid residue of three structural regions—ellipsoidal hollow valve cavity, conical variable diameter section and circular arc turning corner—under different combinations of process parameters; and associating the above data with the corresponding process parameters, flow channel geometric parameters, material parameters and flow channel local morphology tensor field features to form a simulation model training dataset.

[0052] This model structure consists of an input layer, multiple fully connected layers, a physical constraint layer, and an output layer cascaded sequentially. The input layer receives the abrasive concentration, feed rate, polishing pressure, processing time, and flow channel morphology constraint tensors corresponding to three regions: the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner. The fully connected layers extract the high-order coupling relationship between process parameters and morphological features through multi-layer nonlinear transformations. The physical constraint layer embeds the Navier-Stokes equations and the abrasive flow equation as soft constraints to ensure that the model output results conform to the basic physical laws of fluid mechanics and polishing mechanisms. The output layer outputs nine performance indicators for the three regions, including the inner surface roughness of the flow channel, CV value error, and fluid residue. This architecture can fully utilize limited historical data to train a high-precision simulation model while ensuring the physical consistency of the simulation results. Compared with traditional simulation methods based on finite element analysis, it effectively improves the simulation efficiency under multiple working conditions and can capture the nonlinear mapping relationship between process parameters and flow channel morphology features.

[0053] The simulation model training dataset is input into the physical information neural network model to be trained. The deviation between the predicted performance index output by the model and the actual performance index in the training dataset is calculated using forward propagation. A comprehensive loss function is constructed by combining the residual terms of the physical constraint layer. The gradient is calculated layer by layer using the backpropagation algorithm. The network weights are iteratively updated using an optimizer. Training is terminated when the loss function value on the validation set converges, resulting in a pre-trained multi-condition dynamic simulation model. Finally, the process parameter constraint combination to be validated is input into the pre-trained multi-condition dynamic simulation model. The flow channel polishing performance is simulated under multiple preset conditions, including rated conditions and extreme conditions. Rated conditions correspond to the pressure, temperature, and medium conditions when the semiconductor equipment is operating normally. Extreme conditions correspond to boundary conditions such as equipment start-up impact, temperature fluctuation, and changes in medium composition. The simulation model is run under each condition to calculate the simulated values ​​of the inner surface roughness, CV value error, and fluid residue of the flow channel under that condition. The simulation results under each condition are summarized to form an initial simulation result set.

[0054] Step 5.3: Determine whether all performance indicators in the initial simulation result set have reached the preset performance thresholds. If all have reached the thresholds, output the current process parameters and simulation results as the compliance results to obtain the compliance process parameters and compliance simulation verification results. Specifically, this includes comparing the simulated values ​​of the inner surface roughness, CV value error, and fluid residue of the flow channel corresponding to each working condition in the initial simulation result set with the preset performance thresholds one by one. The preset performance thresholds include the target value of the inner surface roughness of the flow channel, the upper limit of the allowable range of CV value error, and the upper limit of the allowable fluid residue. These preset performance thresholds are pre-set according to the requirements of the semiconductor equipment flow channel polishing process. During the comparison, for each working condition, extract the simulated values ​​of the inner surface roughness, CV value error, and fluid residue of the flow channel in the three structural regions (ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner) under that working condition, and compare the above nine performance indicators with the corresponding preset thresholds.

[0055] Then, all performance indicators under all working conditions are summarized and statistically analyzed to determine whether all performance indicators under all working conditions meet the conditions that the simulated value of the inner surface roughness of the flow channel is less than or equal to the target value, the simulated value of the CV value error is less than or equal to the allowable upper limit value, and the simulated value of the fluid residue is less than or equal to the allowable upper limit value. If all performance indicators under all working conditions meet or exceed the preset performance threshold, the currently used flow channel morphology constraint tensor set and the final extrusion honing process parameters are used as the compliant process parameters, and the corresponding initial simulation result set is used as the compliant simulation verification result for output, thus obtaining the compliant process parameters and the compliant simulation verification result.

[0056] Step 5.4: If the initial simulation results set contains performance indicators that do not reach the preset performance threshold, then based on the deviation between the non-compliant indicators and the preset performance threshold, the final extrusion honing process parameters are adjusted in reverse to form optimized process parameters, resulting in iteratively optimized process parameters. Specifically, if the judgment result indicates that the initial simulation results set contains performance indicators that do not reach the preset performance threshold, then the specific type of non-compliant performance indicator is first identified, including one or more of the following: surface roughness of the flow channel, CV value error, or fluid residue. At the same time, the working condition type (rated working condition or extreme working condition) and the structural region to which the non-compliant indicator belongs are identified. Then, the deviation between the non-compliant value and the preset threshold is calculated. This deviation is expressed in absolute value form and is used to quantify the gap between the performance indicator and the target value. The deviation is compared and analyzed with the corresponding region and the constraint value of the corresponding process parameter in the obtained flow channel morphology constraint tensor set. Based on the pre-established sensitivity relationship between process parameters and performance indicators, the type of process parameter that needs to be adjusted is determined, including one or more of the following: abrasive concentration, feed rate, polishing pressure, and processing time. The adjustment direction is determined, i.e., increasing or decreasing the value of the process parameter.

[0057] The pre-established sensitivity relationship between process parameters and performance indicators is a mapping law obtained through regression analysis of historical polishing case data. Specifically, for the surface roughness index of the flow channel, abrasive concentration and polishing pressure have a negative correlation, meaning that increasing abrasive concentration or polishing pressure helps to reduce roughness, while feed rate has a positive correlation, meaning that increasing feed rate will lead to increased roughness. For the CV value error index, polishing pressure and processing time have a negative correlation, meaning that increasing polishing pressure or extending processing time helps to reduce CV value error. For the fluid residue index, abrasive concentration and feed rate have a negative correlation, meaning that increasing abrasive concentration or increasing feed rate helps to reduce fluid residue.

[0058] Based on the matching relationship between the magnitude of the deviation and the constraint values ​​in the flow channel morphology constraint tensor set, the adjustment range is calculated. The matching relationship refers to mapping the deviation to the adjustment amount of the process parameters according to a preset proportional coefficient. This proportional coefficient is set according to the influence intensity in the sensitivity relationship. The larger the deviation, the larger the adjustment range. At the same time, the adjusted process parameter values ​​must not exceed the allowable range of process parameters limited by the corresponding constraint values ​​in the flow channel morphology constraint tensor set, ensuring that the adjusted process parameters are still within a safe and controllable range. The adjustment range is then superimposed on the corresponding process parameters in the corresponding region of the final extrusion honing process parameters in a positive or negative manner to form the optimized process parameters after one correction, resulting in iteratively optimized process parameters.

[0059] Step 5.5 involves re-inputting the iteratively optimized process parameters into the pre-trained multi-condition dynamic simulation model for iterative verification and adjustment until all performance indicators reach the preset performance threshold. This process yields the process parameters that meet the preset performance threshold and the corresponding simulation verification results. Specifically, this includes: re-inputting the obtained iteratively optimized process parameters into the pre-trained multi-condition dynamic simulation model, and performing flow channel polishing performance simulations again under multiple preset conditions to obtain an updated simulation result set; repeating the comparison and judgment operation in step 5.3 on the updated simulation result set, and comparing the surface roughness of the flow channel inner surface, CV value error, and fluid residue in each region under each condition. The simulated values ​​are compared one by one with the preset performance thresholds. If all performance indicators reach the preset performance thresholds, the iteration loop is terminated, and the currently used iteratively optimized process parameters are used as the final process parameters. The updated simulation result set is output as the final simulation verification result. If there are still performance indicators that do not meet the standards, the deviation identification and process parameter adjustment operation in step 5.4 is executed again to generate the process parameters for the next round of iterative optimization. The simulation verification and judgment process in step 5.5 is repeated until all performance indicators under all operating conditions reach the preset performance thresholds. Finally, the process parameters that have reached the preset performance thresholds and the corresponding simulation verification results are obtained.

[0060] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1: The process parameters that reach the preset performance threshold and the corresponding simulation verification results are taken as the core components of the final process scheme to obtain the process parameters and verification data combination to be output. Specifically, this includes: obtaining the process parameters that reach the preset performance threshold and the corresponding simulation verification results, wherein the process parameters that reach the preset performance threshold include the abrasive concentration, feed rate, polishing pressure and processing time values ​​corresponding to the three regions of the ellipsoidal hollow valve cavity, the conical variable diameter section and the arc turning corner, and the simulation verification results include the simulated values ​​of the inner surface roughness of the flow channel, CV value error and fluid residue of each region under rated working conditions and extreme working conditions; organizing these process parameters and simulation verification results according to the region dimension and the working condition dimension to form a structured data combination. This combination is nested and stored with the region as the first level, the working condition as the second level and the process parameters and performance indicators as the third level to obtain the process parameters and verification data combination to be output.

[0061] Step 6.2: Encapsulate the process parameters in the combination of the process parameters to be output and the verification data to form a standard process specification containing the extrusion honing process parameters and their control ranges, resulting in a standard process specification file. Specifically, this includes: encapsulating the process parameters in the combination of the process parameters to be output and the verification data; firstly, extracting the abrasive concentration, feed rate, polishing pressure, and processing time values ​​corresponding to the three regions: the ellipsoidal hollow valve cavity, the conical variable diameter section, and the arc turning corner; setting a control range for each process parameter, with the parameter value as the center and a preset percentage above and below as the allowable fluctuation range; then grouping the above process parameters and their control ranges according to the regions, arranging them in the order of abrasive concentration, feed rate, polishing pressure, and processing time within each group to form standard process specification entries for each region; finally, summarizing the standard process specification entries for the three regions, adding header information including generation time, applicable equipment model, and flow channel identification to form a standard process specification file.

[0062] Step 6.3: The simulation verification results from the combination of process parameters to be output and verification data are categorized and organized according to the performance indicators of inner surface roughness of the flow channel, CV value error, and fluid residue. A compliance verification data table for each performance indicator is generated, resulting in a verification data summary table. Specifically, this includes: categorizing and organizing the simulation verification results from the combination of process parameters to be output and verification data; establishing three independent classification directories based on the three performance indicator types of inner surface roughness of the flow channel, CV value error, and fluid residue; under each classification directory, establishing subdirectories for three regions: ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner; under each region subdirectory, recording the simulation values ​​for that region under the rated operating condition and extreme operating condition, along with the corresponding preset performance thresholds and compliance judgment results; presenting the categorized data in a tabular format according to a three-level structure of performance indicator type, region, and operating condition, generating a compliance verification data table for each performance indicator; merging the compliance verification data tables for the three performance indicators to form a verification data summary table.

[0063] Step 6.4 involves linking and integrating the standard process specification document and the verification data summary table to generate a comprehensive performance verification report containing process parameters, verification results, and performance compliance conclusions. This process includes: linking and integrating the obtained standard process specification document and the obtained verification data summary table; first, adding reference links in the standard process specification document pointing to the corresponding verification data entries in the verification data summary table to ensure that each process parameter in each region is associated with its corresponding performance verification result; then, adding process parameter reference fields in the verification data summary table, pointing back to the corresponding process parameter entries in the standard process specification document, forming a bidirectional association; finally, using the standard process specification document as the main body of the report and the verification data summary table as the appendix, merging them to generate a complete comprehensive performance verification report. This report contains three core content sections: process parameters, verification results, and performance compliance conclusions. The performance compliance conclusions section summarizes and describes the performance indicators compliance status under each region and operating condition, resulting in polishing analysis results including the performance verification report.

[0064] like Figure 2 As shown, embodiments of the present invention also provide a semiconductor device flow channel polishing analysis system based on a channel attention mechanism, comprising: The database construction module is used to build a flow channel-process correlation database, which integrates the geometric parameters, material parameters, extrusion honing process parameters and performance data of semiconductor equipment flow channels, performs normalization preprocessing, and obtains standardized correlation data. The model training and parameter prediction module is used to construct and train a convolutional neural network model containing a channel attention mechanism based on standardized correlation data, so as to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure. The morphological tensor field construction module is used to dynamically extract the topological anchor points of the target flow channel at the ellipsoidal hollow valve cavity, the conical variable diameter section and the arc turning corner, respectively, based on the final extrusion honing process parameters. The flow channel topological anchor points are used to construct a flow channel local morphological tensor field that reflects the local morphological distribution characteristics of the flow channel. The morphological constraint tensor calculation module is used to calculate the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel, and obtain the flow channel morphological constraint tensor set used to constrain the polishing process. The simulation verification and iterative optimization module is used to input the flow channel morphology constraint tensor set and the final extrusion honing process parameters into the pre-trained multi-condition dynamic simulation model. Through simulation verification and iterative optimization, process parameters that reach the preset performance threshold and corresponding simulation verification results are formed. The polishing analysis result generation module generates a final process scheme based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, and obtains polishing analysis results including a performance verification report.

[0065] 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.

[0066] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0067] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0068] 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 semiconductor device flow channel polishing analysis method based on channel attention mechanism, characterized in that, The method includes: Step 1: Construct a flow channel-process correlation database, integrate the geometric parameters, material parameters, extrusion honing process parameters and performance data of semiconductor equipment flow channels, perform normalization preprocessing, and obtain standardized correlation data; Step 2: Based on standardized associated data, construct and train a convolutional neural network model that includes a channel attention mechanism to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure. Step 3: Based on the final extrusion honing process parameters, dynamically extract the topological anchor points of the flow channel at the ellipsoidal hollow valve cavity, the conical variable diameter section, and the circular arc turning corner of the target flow channel. Construct a flow channel local morphology tensor field that reflects the local morphological distribution characteristics of the flow channel from the flow channel topological anchor points. Step 4: Calculate the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel to obtain the flow channel morphological constraint tensor set used to constrain the polishing process. Step 5: Input the flow channel morphology constraint tensor set and the final extrusion honing process parameters into the pre-trained multi-condition dynamic simulation model. After simulation verification and iterative optimization, process parameters that reach the preset performance threshold and corresponding simulation verification results are formed. Step 6: Based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, form the final process scheme and obtain the polishing analysis results including the performance verification report.

2. The semiconductor device flow channel polishing analysis method based on channel attention mechanism according to claim 1, characterized in that, A flow channel-process correlation database is constructed, integrating the geometric parameters, material parameters, extrusion honing process parameters, and performance data of semiconductor equipment flow channels. Normalization preprocessing is then performed to obtain standardized correlation data, including: Collect the geometric and material parameters of the flow channel of the target semiconductor device, construct a flow channel geometry-material data subset, and obtain the first data subset; Collect the extrusion honing process parameters and performance data corresponding to the flow channel geometry-material data subset, construct the process-performance data subset, and obtain the second data subset; The first data subset and the second data subset are associated and mapped according to the flow channel structure characteristics to form the original associated database; The data in the original relational database is preprocessed by normalization to eliminate the differences in dimensions and scales between different data sources, resulting in standardized relational data.

3. The semiconductor device flow channel polishing analysis method based on channel attention mechanism according to claim 2, characterized in that, Based on standardized correlated data, a convolutional neural network model incorporating a channel attention mechanism is constructed and trained to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure, including: The obtained standardized correlation data is used as the basic input data for model training to obtain the training dataset; An initial convolutional neural network model is constructed, and a channel attention mechanism is embedded in the initial convolutional neural network model to form a channel attention convolutional neural network model to be trained. The training dataset is input into the channel attention convolutional neural network model to be trained, and iterative training is performed with the flow channel polishing performance index as the optimization objective to obtain the trained convolutional neural network model. The geometric and material parameters of the target flow channel are input into the trained convolutional neural network model. The trained convolutional neural network model automatically predicts based on the key features weighted by the channel attention mechanism and outputs the final extrusion honing process parameters for the target flow channel structure.

4. The semiconductor device flow channel polishing analysis method based on channel attention mechanism according to claim 3, characterized in that, Based on the final extrusion honing process parameters, flow channel topological anchor points are dynamically extracted at the ellipsoidal hollow valve cavity, conical variable diameter section, and circular arc turning corner of the target flow channel. A flow channel local morphology tensor field reflecting the local morphological distribution characteristics of the flow channel is constructed from these anchor points, including: The final extrusion honing process parameters for the target flow channel structure are obtained as the control benchmark for anchor point extraction, thus obtaining the process parameter constraint set. Based on the set of process parameter constraints, the first set of flow channel topological anchor points is dynamically extracted in the ellipsoidal hollow valve cavity region of the target flow channel according to the curvature change gradient, thus obtaining the valve cavity region topological anchor point set. Based on the process parameter constraint set, in the conical variable diameter section region of the target flow channel, the second set of flow channel topology anchor points is dynamically extracted according to the variable diameter ratio change curve to obtain the topology anchor point set of the variable diameter section region. Based on the set of process parameter constraints, the third set of flow channel topology anchor points is dynamically extracted in the circular arc turning corner region of the target flow channel according to the law of rotation radius variation, so as to obtain the topology anchor point set of the corner region. The topological anchor point set of the valve cavity region, the topological anchor point set of the variable diameter section region, and the topological anchor point set of the corner region are spatially fused to construct a local morphological tensor field of the flow channel that reflects the local morphological distribution characteristics of the flow channel, thus obtaining the local morphological tensor field of the flow channel.

5. The semiconductor device flow channel polishing analysis method based on channel attention mechanism according to claim 4, characterized in that, Calculate the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel to obtain the flow channel morphological constraint tensor set used to constrain the polishing process, including: The constructed local morphological tensor field of the flow channel is obtained as the basic field data for feature quantity calculation, and the morphological tensor field to be analyzed is obtained. Tensor decomposition is performed on the morphological tensor field to be analyzed, and the principal eigenvalues ​​and principal eigendirections reflecting the local geometry of the flow channel are extracted to obtain the local morphological principal feature set. Based on the local morphological main feature set, the curvature tensor feature of the ellipsoidal hollow valve cavity region, the taper change rate feature of the conical variable diameter section region, and the rotation curvature feature of the circular arc turning corner region are calculated respectively to obtain the regionalized morphological feature set. By fusing the regionalized morphological feature sets, a comprehensive feature vector is constructed to characterize the overall morphological constraint capability of the flow channel, thus obtaining the comprehensive morphological constraint feature vector. By mapping the comprehensive morphological constraint feature vector to the polishing process parameter space, a flow channel morphological constraint tensor set that can quantitatively constrain the extrusion honing process parameters is formed, thus obtaining the flow channel morphological constraint tensor set used to constrain the polishing process.

6. The semiconductor device flow channel polishing analysis method based on channel attention mechanism according to claim 5, characterized in that, The process parameters and corresponding simulation verification results that achieve the preset performance threshold through simulation verification and iterative optimization include: The flow channel morphology constraint tensor set and the final extrusion honing process parameters are used as joint input data for simulation verification and optimization to obtain the process parameter constraint combination to be verified. The process parameter constraints to be verified are input into the pre-trained multi-condition dynamic simulation model. The flow channel polishing performance is simulated under multiple preset conditions to obtain the initial simulation result set of the flow channel inner surface roughness, CV value error and fluid residue corresponding to each condition. Determine whether all performance indicators in the initial simulation result set have reached the preset performance threshold. If all have reached the threshold, output the current process parameters and simulation results as the compliance results to obtain the compliance process parameters and compliance simulation verification results. If there are performance indicators in the initial simulation results set that do not reach the preset performance threshold, then the final extrusion honing process parameters are adjusted in reverse according to the deviation between the non-compliant indicators and the preset performance threshold to form optimized process parameters, and thus obtain iteratively optimized process parameters. The iteratively optimized process parameters are re-inputted into the pre-trained multi-condition dynamic simulation model for iterative verification and adjustment until all performance indicators reach the preset performance threshold, thus obtaining the process parameters that have reached the preset performance threshold and the corresponding simulation verification results.

7. The semiconductor device flow channel polishing analysis method based on channel attention mechanism according to claim 6, characterized in that, Based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, a final process scheme is formed, resulting in polishing analysis results including a performance verification report, including: The process parameters that reach the preset performance threshold and the corresponding simulation verification results are used as the core components of the final process scheme to obtain the combination of process parameters and verification data to be output. The process parameters to be output are combined with the process parameters in the verification data and encapsulated to form a standard process specification containing the extrusion honing process parameters and their control range, thus obtaining a standard process specification file. The simulation verification results in the combination of the process parameters to be output and the verification data are classified and organized according to the performance indicators of the inner surface roughness of the flow channel, CV value error and fluid residue, and a standard verification data table for each performance indicator is generated to obtain a verification data summary table. By linking and integrating the standard process specification documents with the verification data summary table, a comprehensive performance verification report is generated, which includes process parameters, verification results, and performance compliance conclusions. The polishing analysis results containing the performance verification report are then obtained.

8. A semiconductor device flow channel polishing analysis system based on channel attention mechanism, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The database construction module is used to build a flow channel-process correlation database, which integrates the geometric parameters, material parameters, extrusion honing process parameters and performance data of semiconductor equipment flow channels, performs normalization preprocessing, and obtains standardized correlation data. The model training and parameter prediction module is used to construct and train a convolutional neural network model containing a channel attention mechanism based on standardized correlation data, so as to obtain a trained convolutional neural network model. The trained convolutional neural network model outputs the final extrusion honing process parameters for the target flow channel structure. The morphological tensor field construction module is used to dynamically extract the topological anchor points of the target flow channel at the ellipsoidal hollow valve cavity, the conical variable diameter section and the arc turning corner, respectively, based on the final extrusion honing process parameters. The flow channel topological anchor points are used to construct a flow channel local morphological tensor field that reflects the local morphological distribution characteristics of the flow channel. The morphological constraint tensor calculation module is used to calculate the morphological tensor field characteristic quantities of the local morphological tensor field of the flow channel, and obtain the flow channel morphological constraint tensor set used to constrain the polishing process. The simulation verification and iterative optimization module is used to input the flow channel morphology constraint tensor set and the final extrusion honing process parameters into the pre-trained multi-condition dynamic simulation model. Through simulation verification and iterative optimization, process parameters that reach the preset performance threshold and corresponding simulation verification results are formed. The polishing analysis result generation module generates a final process scheme based on the process parameters that reach the preset performance threshold and the corresponding simulation verification results, and obtains polishing analysis results including a performance verification report.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images