A deep learning-based wafer-like component performance rapid prediction method

Abstract

The present application relates to the field of artificial intelligence, in particular to a kind of sheet component performance fast prediction method based on deep learning, collect multi-working condition simulation data to generate training sample, construct and encode grid topology structure to extract multidimensional feature, after characteristic fusion and self-attention mechanism processing, input perception machine to predict stress and evaluate error.In the present application, structured training sample set is constructed by introducing simulation information, combined with the geometric structure feature set formed by node space coordinates, boundary constraints and connection relationship, so that the mutual position and constraint conditions between nodes are completely expressed in the graph structure, through node-level feature extraction and feature fusion processing, the deep embedding of node geometric layout and boundary mutual relationship is realized, the learnable expression of stress evolution path in spatial structure is established through local subgraph and context analysis, multi-layer feature aggregation and attention mechanism are introduced in the node graph embedding process, and the feature response expression of key area is strengthened.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for quickly predicting the performance of thin-film components based on deep learning. Background Art

[0002] The field of artificial intelligence technology involves systems, algorithms, methods and technologies that simulate, extend or expand human intelligence, including machine learning, deep learning, natural language processing, computer vision, intelligent control, etc.

[0003] Among them, the performance prediction of thin-film components refers to components with small thickness, large plane size or complex geometric features such as gaskets, sealing rings, printed circuit boards (PCBs), silicon steel sheets, and thin-film sensors. It usually requires multiple rounds of physical performance tests or relies on finite element simulation methods based on refined meshes and high-fidelity material models to determine their performance under different loads, vibrations, thermal cycles, stress shocks and other working conditions.

[0004] Existing technologies rely on multiple rounds of physical performance tests and high-precision finite element simulations to evaluate the performance of thin-film components under thermal cycling and multi-load conditions. However, due to the requirements of complex geometric structures for mesh refinement, the model construction cycle is easily prolonged during the construction process. In particular, when boundary conditions and loading methods vary, modeling inputs need to be frequently modified and re-simulated, resulting in low modeling efficiency and high simulation costs. In practical applications, tracking the changing trend of stress response in a local area within the structure is difficult to quickly reflect in the overall prediction model, resulting in insufficient accuracy of the prediction results in boundary-sensitive areas, further affecting the reliability and working condition coverage of the overall performance evaluation. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the prior art and to propose a method for quickly predicting the performance of thin-film components based on deep learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for rapidly predicting the performance of thin-film components based on deep learning, comprising the following steps:

[0007] S1: Collect simulation information of silicon steel sheets under various load and thermal cycle conditions to obtain a structured training sample set;

[0008] S2: establishing a grid topology structure of the silicon steel sheet according to the silicon steel sheet simulation information in the structured training sample set, encoding the grid topology structure, and outputting the encoding input elements;

[0009] S3: Inputting the coding input elements into a deep neural network, extracting input information of the geometric positions of silicon steel sheet nodes and the relationship between the boundaries from the silicon steel sheet grid topology structure, and outputting Trunk network coding features and Branch network coding features of the input information through network coding;

[0010] S4: Fusing the Trunk network coding features and the Branch network coding features, enhancing local features through the self-attention mechanism, and obtaining a unified fusion representation vector;

[0011] S5: Inputting the unified fusion representation vector into a multi-layer perceptron to predict the stress of the silicon steel sheet, calling the simulation information in the structured training sample set to evaluate the prediction error, and obtaining the silicon steel sheet performance prediction result.

[0012] As a further solution of the present invention, the structured training sample set includes load type, thermal cycle conditions, and node response data; the encoding input elements are specifically grid connection relationships, topological structure labels, and geometric positioning parameters; the Trunk network encoding features are specifically node spatial distribution patterns and global structural features; the Branch network encoding features are specifically boundary adjacency relationship features and local geometric features; the unified fusion representation vector includes local enhancement features, multi-layer feature aggregation results, and attention distribution weights; the silicon steel sheet performance prediction results include stress numerical distribution, prediction deviation indicators, and performance evaluation outputs.

[0013] As a further solution of the present invention, the steps of obtaining the structured training sample set are specifically as follows:

[0014] S111: Collect simulation information of silicon steel sheets under various load and thermal cycle conditions, including node space coordinates, unit connection relationships, and boundary constraint parameters. The node space coordinates correspond to the grid node positions in the silicon steel sheet geometry, the unit connection relationships correspond to the combination methods in the silicon steel sheet geometry, and the boundary constraint parameters correspond to the force and restricted area settings in the silicon steel sheet loading condition properties. Extract simulation input structure information according to the loading number and generate a node configuration and constraint information set.

[0015] S112: Based on the node configuration and constraint information set, extract the principal stress direction and principal stress tensor of the corresponding node in each thermal cycle stage, arrange the principal stress direction vectors of each stage according to the thermal loading sequence, calculate the angle change value of the direction data of the same node between consecutive stages, and record the direction change path;

[0016] S113: Integrate the node space coordinates, unit connection relationship, principal stress tensor, direction change path, boundary constraint parameters and loading number corresponding to each node to generate a structured training sample set.

[0017] As a further solution of the present invention, the step of obtaining the coding input element is specifically as follows:

[0018] S211: obtaining all node numbers and corresponding coordinate positions according to the silicon steel sheet simulation information in the structured training sample set, identifying connection paths between nodes based on unit connection relationships, and constructing a silicon steel sheet grid topology structure;

[0019] S212: Based on the silicon steel sheet grid topology, extract the spatial direction information of each node within the local topology range, calculate the direction vector and Euclidean length of the edge between adjacent nodes, obtain the point normal vector in units of nodes as the point feature, use the edge direction and edge length as the edge feature, call the fast point cloud normal vector estimation algorithm to correct the point normal vector, and generate a geometric structure feature set;

[0020] S213: Integrate the geometric structure feature set into a model input format in the form of a graph structure expression, and output the encoded input elements.

[0021] As a further solution of the present invention, the steps for obtaining the Trunk network coding features and the Branch network coding features are specifically as follows:

[0022] S311: Inputting the node spatial coordinates in the encoded input elements into the Trunk network of the deep neural network, where the node spatial coordinates correspond to the geometric positions of the nodes in the silicon steel sheet grid topology structure, arranging the node numbers in the order of the graph structure while keeping the connection relationship unchanged, and using the relative distribution relationship between the nodes to capture the arrangement characteristics in the structural topology, generating the Trunk network geometric layout features;

[0023] S312: Inputting the boundary constraint parameters, point normal vectors, edge directions, and edge lengths in the encoded input elements into the Branch network of the deep neural network, wherein the boundary constraint parameters correspond to the boundary interaction conditions of the silicon steel sheet structure, the point normal vectors represent the node direction attributes, and the edge directions and edge lengths reflect the directionality of the structural connection, thereby generating the Branch network boundary response characteristics;

[0024] S313: Arrange the branch network boundary interaction features and the trunk network geometric position features into a multi-dimensional coding vector corresponding to the structural position index to obtain the trunk network coding features and the branch network coding features.

[0025] As a further solution of the present invention, the step of obtaining the unified fusion representation vector is specifically as follows:

[0026] S411: Fuse the Trunk network coding features and Branch network coding features, correspond the node geometric position features in the Trunk network coding features and the boundary interaction features in the Branch network coding features according to the node index, perform splicing combination on the two types of feature vectors to generate a node-level joint coding feature sequence;

[0027] S412: Input the node-level joint coding feature sequence into a cross-attention mechanism, use the position index of the node in the topology structure as a query item to construct the mutual attention weight relationship between the joint coding features, and generate a cross-matching feature representation vector;

[0028] S413: According to the cross-matching feature representation vector, construct a local subgraph using the connection relationship of adjacent nodes in the graph, analyze the context features of each node according to the weight value of the connected nodes, and generate a unified fusion representation vector.

[0029] As a further scheme of the present application, the performance prediction result obtaining step is specifically:

[0030] S511: Input the unified fusion representation vector into a multilayer perception machine, the unified fusion representation vector corresponds to the graph embedding expression of each node in the silicon steel sheet grid topology structure, input the continuous network layer of the perception machine according to the node index order, perform layer-by-layer vector mapping and nonlinear activation, and generate a node stress prediction result;

[0031] S512: According to the node stress prediction result, call the simulation information in the structured training sample set, match the corresponding target stress value according to the node index, establish the mapping and comparison relationship between the predicted value and the target value, calculate the residual term and the loss value in the root mean square error function, and generate a silicon steel sheet performance prediction result.

[0032] Compared with the prior art, the present application has the following advantages and positive effects:

[0033] In the present application, the structured training sample set is constructed by introducing simulation information, and the geometric structure feature set formed by combining node spatial coordinates, boundary constraints and connection relationships, so that the mutual position and constraint conditions between nodes are completely expressed in the graph structure. Through node-level feature extraction and feature cross-fusion processing, the deep embedding of node geometric layout and boundary mutual relationship is realized. Through local subgraph and context analysis, the learnable expression of stress evolution path in the spatial structure is established. In the node graph embedding process, the multilayer feature aggregation and attention mechanism are introduced to strengthen the feature response expression of the key area, thereby effectively predicting the structural stress response distribution under different thermal cycles and load conditions, reducing the dependence on refined grid modeling and high-fidelity physical testing, and improving the response speed and application range of performance prediction. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 It is a schematic diagram of the main steps of the present invention;

[0035] Figure 2 This is a flow chart of step S1 of the present invention;

[0036] Figure 3 This is a flow chart of step S2 of the present invention;

[0037] Figure 4 This is a flow chart of step S3 of the present invention;

[0038] Figure 5 This is a flow chart of step S4 of the present invention;

[0039] Figure 6 This is a flow chart of step S5 of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0041] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0042] See also Figure 1 The present invention provides a technical solution: a method for quickly predicting the performance of thin-film components based on deep learning, comprising the following steps:

[0043] S1: Collect simulation information of silicon steel sheets under various load and thermal cycle conditions to obtain a structured training sample set;

[0044] S2: Based on the simulation information of the silicon steel sheet in the structured training sample set, a grid topology structure of the silicon steel sheet is established, the grid topology structure is encoded, and the encoded input elements are output;

[0045] S3: Input the coded input elements into the deep neural network, extract the input information of the geometric position of the silicon steel sheet nodes and the relationship between the boundaries from the silicon steel sheet grid topology structure, and output the trunk network coding features and branch network coding features of the input information through network coding.

[0046] S4: Fuse the trunk network coding features with the branch network coding features, enhance the local features through the self-attention mechanism, and obtain a unified fusion representation vector;

[0047] S5: Input the unified fusion representation vector into the multi-layer perceptron to predict the stress of the silicon steel sheet, call the simulation information in the structured training sample set to evaluate the prediction error, and obtain the performance prediction result of the silicon steel sheet;

[0048] The structured training sample set includes load type, thermal cycle conditions, and node response data. The encoding input elements are specifically grid connection relationships, topological structure labels, and geometric positioning parameters. The trunk network encoding features are specifically node spatial distribution patterns and global structural features. The branch network encoding features are specifically boundary adjacency relationship features and local geometric features. The unified fusion representation vector includes local enhancement features, multi-layer feature aggregation results, and attention distribution weights. The silicon steel sheet performance prediction results include stress numerical distribution, prediction deviation indicators, and performance evaluation outputs.

[0049] See also Figure 2 , the steps for obtaining the structured training sample set are as follows:

[0050] S111: Collect simulation information of silicon steel sheets under various load and thermal cycle conditions, including node space coordinates, unit connection relationships, and boundary constraint parameters. The node space coordinates correspond to the grid node positions in the silicon steel sheet geometry, the unit connection relationships correspond to the combination methods in the silicon steel sheet geometry, and the boundary constraint parameters correspond to the force and restricted area settings in the silicon steel sheet loading condition properties. Extract simulation input structure information according to the loading number and generate a node configuration and constraint information set.

[0051] To collect simulation information of silicon steel sheets under various load and thermal cycle conditions, it is necessary to first establish a geometric model of the silicon steel sheet in a simulation platform (such as Abaqus, ANSYS), set appropriate meshing parameters to generate node space coordinate information, such as refining the mesh in the thickness direction of the silicon steel sheet to facilitate subsequent stress extraction, and define the unit connection relationship in the geometric model, clarify the connection topology of each unit with the surrounding units, and establish a combination method through unit type selection (such as linear hexahedron or tetrahedron). In terms of load application, boundary constraint parameters should be set separately for the actual working conditions corresponding to different loading numbers, such as applying displacement constraints on the edge to simulate the clamping state, and in the center area, Apply loads to simulate tension or compression. For thermal cycle loading, input heat flux or temperature change information in numerical order, and calculate thermal stress evolution in combination with the thermal-mechanical coupling analysis module. The acquisition process includes exporting the three-dimensional coordinate value of each node, such as "x direction is a certain value, y direction is a certain value, z direction is a certain value", and the unit connection relationship is output in the form of element number index node combination, such as "unit 1 connects node 1, node 2, node 3, node 4", etc. Boundary constraints are recorded through command input to apply constraints to each node, such as "node 5 applies a fixed constraint in the x direction", and the structure information table is organized according to the loading number. The above information is integrated to construct the node configuration and constraint information set.

[0052] S112: Based on the node configuration and constraint information set, extract the principal stress direction and principal stress tensor of the corresponding node in each thermal cycle stage, arrange the principal stress direction vectors of each stage according to the thermal loading sequence, calculate the angle change value of the direction data of the same node between consecutive stages, and record the direction change path;

[0053] To extract the principal stress directions and principal stress tensors of nodes in the thermal cycle stage based on the node configuration and constraint information set, it is necessary to first set the thermal cycle conditions, including initial temperature, heating rate, constant temperature holding time and temperature gradient in the cooling stage, etc., define these stages as load steps in the simulation software, and export the stress tensor value of each node at each time step after running. According to the material constitutive relationship, the stress tensor is converted into principal stress value and principal stress direction. The principal stress tensor extraction method can obtain the principal stress value and principal direction vector of each node in a certain stage through the software post-processing function, and then arrange the principal stress direction vectors of the same node in each thermal stage according to the loading order, and compare the principal stress directions of adjacent stages. For example, the direction in the first stage is vector A, and the direction in the second stage is vector B. The angle is the arccosine angle between vector A and vector B. Specifically, the direction change angle = the dot product of direction vector A and direction vector B divided by the arccosine function value after the product of the two vector moduli is calculated. The adjacent angle values ​​in each stage are recorded as the direction change path. For the numerical setting of the angle, 0 to 90 degrees can be divided into three intervals: low (less than 30 degrees), medium (30 to 60 degrees) and high (greater than 60 degrees). Then, whether the direction change is drastic is judged. By recording the direction change path of the node in the entire thermal cycle process, a tensor direction evolution sequence for subsequent training samples is formed.

[0054] S113: Integrate the node space coordinates, element connection relationship, principal stress tensor, direction change path, boundary constraint parameters and loading number corresponding to each node to generate a structured training sample set;

[0055] Integrate the node space coordinates, unit connection relationship, principal stress tensor, direction change path, boundary constraint parameters and loading number to generate a structured training sample set. It is necessary to unify the data format first, record the node space coordinates in the form of a three-dimensional array, such as "node number corresponding to x, y, z position", and form an index table of unit connection relationship, such as "unit number and list containing node number". The principal stress tensor value must include the three principal stress values ​​(maximum, minimum and intermediate values) of the node in each thermal cycle stage and their corresponding direction vectors. The direction change path is organized in the form of stage number and direction angle pairs, such as "stage 1-2 angle xx degrees", and the boundary constraint parameters are recorded by recording the node constraints. Dimensions such as "node 1 fixed x direction, free y and z directions" are supplemented, and the loading number is used to identify the data source condition. All information is organized in a dictionary nested structure to form a training sample dictionary list. A set of data samples is generated for each node. The sample contains spatial position information, connection topology, principal stress value of the thermal cycle stage, direction evolution sequence and boundary constraint set. For example, the sample structure of a node can be expressed as: "The position is x=10, y=0, z=0, the connection units are elements 12 and 15, the principal stress of thermal stage 1 is xx MPa, the direction is (a, b, c), the angle is xx degrees, and the number is condition A", thus forming a complete structured data sample.

[0056] See also Figure 3 , the steps for obtaining the coded input elements are as follows:

[0057] S211: According to the silicon steel sheet simulation information in the structured training sample set, all node numbers and corresponding coordinate positions are obtained, connection paths between nodes are identified based on the unit connection relationship, and a silicon steel sheet grid topology structure is constructed;

[0058] Read each record in the structured training sample set, obtain the node number and coordinate position, extract the spatial position value of each node, including the x-direction coordinate, y-direction coordinate, and z-direction coordinate, and establish a spatial coordinate index table with the node number as the index. Then traverse all unit connection relationships, list the multiple node numbers contained in each unit, and combine these node numbers in turn to form paired connection paths. Perform the combination on all units to build a global connection edge set, and then query the spatial coordinate position of each pair of connection edges according to the node number in each pair of connection edges. Establish a node connection path mapping table, group all connection edges according to the node number, and each node corresponds to multiple connection edges to form a preliminary grid topology relationship. For example, if node 1 is connected to nodes 2, 3, and 4, then node 1 has three connection edges, and connections are also established between node 2 and node 3, and node 3 and node 4, thus forming a closed topological structure. Each edge in the topological structure represents the connection relationship between nodes. Combined with the node coordinate information, the grid topology structure of the silicon steel sheet is formed.

[0059] S212: Based on the silicon steel sheet grid topology, the spatial direction information of each node within the local topology range is extracted, the direction vector and Euclidean length of the edge between adjacent nodes are calculated, the point normal vector in the node unit is obtained as the point feature, the edge direction and edge length are used as the edge feature, and the fast point cloud normal vector estimation algorithm is used to correct the point normal vector to generate a geometric structure feature set;

[0060] The local topology information extraction operation is performed on each node in the topology structure, all adjacent node numbers are obtained from the connection path of the node, the spatial coordinate values of the adjacent nodes are found out, the adjacent node x direction coordinate-current node x direction coordinate=x direction component, the adjacent node y direction coordinate-current node y direction coordinate=y direction component, and the adjacent node z direction coordinate-current node z direction coordinate=z direction component are used to form the direction vector of the edge, the length of the direction vector is obtained by x direction component x x direction component+y direction component x y direction component+z direction component x z direction component=edge length square, and the direction vector and the edge length together form the edge feature. After combining the direction vectors and the edge length of multiple connection edges of each node, the geometric input data of the node in the local topology range is constructed, the initial point normal vector can be represented by the unit vector obtained by adding multiple direction vectors, or the optimal estimated direction can be calculated by minimizing the angle between the direction vectors, the initial value is corrected by calling the point cloud normal vector estimation algorithm, the angle between all direction vectors and the initial normal vector is counted during the calculation, the minimum deviation direction fitting is performed using the angle change value, the angle is obtained by taking the inverse cosine function after the angle cosine value is obtained by direction vector·normal vector÷(direction vector module x normal vector module)=angle cosine value, and the final point normal vector is obtained after all adjacent direction vectors are summarized, which is used as the point feature of the node, the direction vector and the edge length of the connection edge are used as the edge feature, and the point feature and the edge feature of all nodes are combined to form a geometric structure feature set.

[0061] S213: integrate the geometric structure feature set into a model input format in the form of a graph structure expression, and output an encoded input element;

[0062] According to the characteristic information of all nodes and edges, they are organized into a model input format in the form of a graph structure expression. Each node forms an input data unit, which records the node number and the three directional components of the point normal vector to form a node feature vector. The node feature vector format is [point normal vector x component, point normal vector y component, point normal vector z component]. Each edge records the starting node number, the ending node number, the direction vector and the edge length. The three components of the direction vector are the ending node x coordinate - the starting node x coordinate = x direction component, the ending node y coordinate - the starting node y coordinate = y direction component, the ending node z coordinate - the starting node z coordinate = z direction component, and the edge length It is x-direction component × x-direction component + y-direction component × y-direction component + z-direction component × z-direction component = the square of the length of the edge. The node features and edge features are organized as an input matrix, and each row represents a node or an edge. In addition, the adjacency matrix A is constructed, A(i,j)=1 indicates that there is a connection between node i and node j, and A(i,j)=0 indicates that there is no connection between nodes i and j. The three structures together constitute the graph structure input file, which is used to describe the complete silicon steel sheet geometric feature graphic model input format, ensuring that each node and its connecting edges are accurately expressed in the input. The input elements after encoding are the collection of node feature matrix, edge feature matrix and adjacency matrix.

[0063] See also Figure 4 The steps for obtaining the Trunk network coding features and the Branch network coding features are as follows:

[0064] S311: Inputting the node spatial coordinates in the encoded input elements into the Trunk network of the deep neural network, where the node spatial coordinates correspond to the geometric positions of the nodes in the silicon steel sheet grid topology structure, arranging the node numbers in the order of the graph structure while keeping the connection relationship unchanged, and using the relative distribution relationship between the nodes to capture the arrangement characteristics in the structural topology, generating the geometric layout features of the Trunk network;

[0065] The input of the node space coordinates in the encoding input element into the Trunk network of the deep neural network needs to read the x-direction coordinate, y-direction coordinate and z-direction coordinate of each node first, and arrange them according to the node number order defined in the graph structure to construct a three-dimensional coordinate input sequence, ensure that the positions of adjacent nodes in the sequence remain unchanged in the connection state of the grid structure, each group of coordinates consists of three numerical values, for example, [x-direction coordinate, y-direction coordinate, z-direction coordinate] as a group of input items, then input the sequence into the Trunk network for feature extraction, the neural network traverses the input sequence in turn and analyzes the arrangement according to the position relationship between each node and its adjacent nodes, the relative distribution relationship is obtained by calculating the coordinate difference between node i and node j, the coordinate difference calculation process is x-direction coordinate of node i-x-direction coordinate of node j=x-direction relative position, y-direction coordinate-y-direction coordinate=y-direction relative position, z-direction coordinate-z-direction coordinate=z-direction relative position, the Trunk network takes such difference values as part of the input weight inside the network, which is used to establish the position mapping weight relationship between adjacent nodes, this weight relationship reflects the continuous mode of geometric arrangement characteristics, and finally the Trunk network generates a set of geometric layout feature vectors based on all node space relationships, each row in the vector set corresponds to the geometric representation characteristics of a node in its graph structure.

[0066] S312: input the boundary constraint parameter, point normal vector and edge direction and edge length in the encoding input element into the Branch network of the deep neural network, the boundary constraint parameter corresponds to the boundary interaction condition of the silicon steel sheet structure, the point normal vector represents the node direction attribute, and the edge direction and edge length reflect the structural connection directionality to generate a Branch network boundary response feature;

[0067] The boundary constraint parameters, point normal vectors, edge directions, and edge lengths in the encoded input elements are input into the Branch network of the deep neural network. The boundary constraint information, point normal vector values, connected edge direction vectors, and edge length values ​​of each node need to be extracted in turn. The boundary constraint information is represented by a three-dimensional restriction mark, which records whether the x direction is restricted, whether the y direction is restricted, and whether the z direction is restricted. For example, the boundary constraint of node k is represented as [whether the x direction is restricted, whether the y direction is restricted, whether the z direction is restricted]. The point normal vector is composed of three direction components [normal vector x component, normal vector y component, normal vector z component]. Each edge direction vector is obtained by calculating the difference in coordinates of adjacent nodes, such as the x coordinate of the starting point of the edge - the x coordinate of the end point = the x direction vector Component, the edge length is calculated by the three directional components of the direction vector, and the formula is x-direction vector component × x-direction vector component + y-direction vector component × y-direction vector component + z-direction vector component × z-direction vector component = the square of the length of the edge. The Branch network receives all the above eigenvalues, encodes them into a set of composite input vectors, and processes each set of input vectors in parallel to extract the response change characteristics between each physical property and topological property. For example, whether the boundary state change of the node affects the directional characteristics or normal feature distribution of the edge, and interactively maps these combined feature vectors to generate the boundary response feature set output by the Branch network. Each row corresponds to a node and its boundary-related encoding representation vector.

[0068] S313: Arrange the branch network boundary interaction features and the trunk network geometric position features into a multi-dimensional coding vector corresponding to the structural position index to obtain the trunk network coding features and the branch network coding features;

[0069] The boundary interaction features output by the Branch network and the geometric position features output by the Trunk network are combined and arranged into a multi-dimensional coding vector corresponding to the structural position index. First, the geometric feature vector output by the Trunk network and the boundary feature vector output by the Branch network are matched one by one with node number as the index. The corresponding operation needs to ensure that the node index sequence is consistent. For example, the node m corresponds to the feature vector [geometric value 1, geometric value 2, geometric value 3] in the Trunk network, and corresponds to the feature vector [boundary value 1, boundary value 2, boundary value 3] in the Branch network. The two vectors are spliced to form [geometric value 1, geometric value 2, geometric value 3, boundary value 1, boundary value 2, boundary value 3] as the final multi-dimensional coding vector of node m. The length of the spliced vector is equal to the dimension of the Trunk vector plus the dimension of the Branch vector. The same operation is performed on all nodes to generate a list of multi-dimensional coding vectors, which is sorted according to the node number to form a complete graph structure network input matrix. Each row in the matrix represents the structural position index of a node and its joint coding result, which serves as the basic coding feature input for subsequent structural stress prediction, thermal deformation mapping and other tasks.

[0070] Please refer to Figure 5 The step of obtaining a unified fusion representation vector is specifically:

[0071] S411: Fuse the Trunk network coding features and the Branch network coding features. The node geometric position features in the Trunk network coding features and the boundary interaction features in the Branch network coding features are matched according to the node index. The two types of feature vectors are spliced and combined to generate a node-level joint coding feature sequence.

[0072] When fusing trunk and branch network coding features, the geometric position feature vector corresponding to each node in the trunk network must first be extracted. This vector is generated by processing the node's spatial coordinate relationship in a deep neural network. Simultaneously, the boundary interaction feature vector generated by the node in the branch network is extracted. This vector is derived from the parallel processing of parameters such as the point normal vector, edge direction, edge length, and boundary constraints. The order of concatenation of the two vectors must be strictly indexed according to the node number. Specifically, for each node number k, the corresponding geometric position feature vector is extracted from the trunk network. This vector includes the node's spatial coordinate features (e.g., geometric value in the x-direction, geometric value in the y-direction, and geometric value in the z-direction). Simultaneously, the boundary interaction feature vector corresponding to the node is extracted from the branch network. This vector includes the boundary constraints, point normal vector components, and local response features related to edge direction and length (e.g., boundary interaction value 1, boundary interaction value 2, and boundary interaction value 3). Subsequently, these two sets of vectors are strictly aligned according to the node number and concatenated to form the joint coding feature vector for node k. For example, if the trunk network coding feature of a node is [geometric position value 1, geometric position value 2, geometric position value 3], and its branch network coding feature is [boundary interaction value 1, boundary interaction value 2, boundary interaction value 3], then the final concatenated joint feature vector is [geometric position value 1, geometric position value 2, geometric position value 3, boundary interaction value 1, boundary interaction value 2, boundary interaction value 3]. By performing the same concatenation operation on all nodes in sequence, a complete node-level joint coding feature sequence is formed. Each item in the sequence corresponds to a node number, completely preserving its position order in the topological structure and not disrupting the original structural connection path.

[0073] S412: Input the node-level joint encoding feature sequence into the cross attention mechanism, use the node position index in the topological structure as the query item, build the mutual attention weight relationship between the joint encoding features, and generate a cross-matching feature representation vector;

[0074] In the finite element mesh structure modeling of silicon steel sheets, nodes are geometric coordinate points obtained during the meshing process. Each node has a definite position in three-dimensional space and simultaneously carries local physical quantities such as boundary conditions, principal stress states, and normal directions. Each node not only represents a part of the geometric structure but also reflects its local response under coupled thermal and magnetic loading scenarios. Therefore, when modeling the information of these nodes in a deep graph network, it is necessary to consider the combined input of their spatial and boundary characteristics.

[0075] When executing the cross attention mechanism, the node joint encoding feature matrix is ​​first linearly transformed. Assume that there are a total of nodes, and the joint encoding feature dimension of each node is , these feature matrices are expressed as: , ,in, : No. The joint encoding feature vector of nodes; : the total number of nodes in the silicon steel sheet grid; : The dimension of each node vector, including spatial position, normal direction, boundary constraints, connection edge direction and length, etc.

[0076] The three main input matrices in the crisscross attention mechanism are constructed from this matrix:

[0077] , , ;

[0078] in, : Query matrix (Query), which indicates the direction of attention of each node to other nodes’ information; : Key matrix (Key), which represents the features exposed to the outside by each node; : Value matrix (Value), which represents the information that can be aggregated for each node; : is a learnable weight matrix; : The dimension of the attention map subspace; : Node feature matrix, which is arranged in the order of node numbers when input, corresponding to the consistency of node numbers in the silicon steel sheet topology structure.

[0079] For any node in the silicon steel sheet , define its adjacent node set as , that is, with the node The numbered set of all other nodes that have a connection relationship comes from the node coincidence connection between grid units. Under this adjacency structure, the node For any of its adjacent nodes The attention weight The calculation is as follows:

[0080] ;

[0081] in, :matrix Middle Row, corresponding node The query vector of :matrix Middle The transpose of the row vector, corresponding to the node The key vector of : Indicates a node With node The vector dot product of , measures the similarity of their features; : exponential function; :For all nodes Adjacent nodes Sum, used to normalize weights; :node For Node The attention weight of node Feature pair nodes degree of impact.

[0082] Then calculate the node The output feature representation , which is a node Neighbor node value vector The weighted sum of:

[0083] ;

[0084] in, :node The cross attention output vector of : Attention weight from the previous step; :node The value vector of , representing the node The coded information content; :For nodes All neighbor nodes of Performs a weighted sum operation.

[0085] The essence of this process is to allow each node to adaptively aggregate local structural information based on the joint characteristics of its adjacent nodes on the basis of maintaining the topological structure of the silicon steel sheet grid, and finally form a new coding expression This expression reflects the local and contextual state linkage behavior of a node under various thermal cycles and load combinations. After all nodes complete this process, a set of output vectors is formed, which will be used as input for subsequent graph convolution propagation, structural response prediction, or damage identification models.

[0086] Assume that three nodes are extracted from the local area of ​​the silicon steel sheet structure, numbered as node 1, node 2, and node 3. These nodes come from the three-dimensional finite element mesh established during the simulation of the silicon steel sheet. There is a topological connection between the nodes, as if they belong to the same mesh unit or are connected by edges. The joint encoding feature dimension of each node is set to , each dimension represents: the first dimension: compressed representation of the node space position (such as x-coordinate projection); the second dimension: normal directional component; the third dimension: quantized value of boundary restriction strength; the fourth dimension: response factor related to the direction of the adjacent edge.

[0087] Assume the eigenvectors are as follows: , , ;

[0088] Assume that node 1 is connected to node 2 and node 3 (i.e., they are its adjacent nodes), and use these vectors directly in the attention mechanism.

[0089] To facilitate the demonstration of calculations, assume that the projection matrix of query, key, and value is the identity matrix, that is:

[0090] ,but: , , ;

[0091] Right now: , , .

[0092] Calculate the attention score (dot product):

[0093] ;

[0094] .

[0095] Calculate attention weights:

[0096] ;

[0097] ;

[0098] Compute node 1 output vector (weighted sum):

[0099] ;

[0100] Calculate the weighted values ​​of two vectors separately:

[0101] ;

[0102] ;

[0103] Add vectors element-wise:

[0104] ;

[0105] The final cross-attention output vector of node 1 is: , where: the first dimension reflects the structural position adjustment characteristics of node 1 after integrating the positions of surrounding nodes; the second dimension reflects the directionality of the integrated normal response; the third dimension reflects the integrated boundary strength characteristics; and the fourth dimension reflects the response adjustment amount influenced by the direction of the topological connection edge. This vector serves as the basic input for node 1's further propagation in the graph neural network, reflecting the joint physical state formed by it and the surrounding nodes in the thermal, mechanical, and magnetic multi-field coupled response environment of the silicon steel sheet.

[0106] First, three vector sets are generated based on the encoded features of the nodes in the silicon steel sheet structure, representing the node's attention intention to external information, the feature content exposed by the node itself, and the actual information carried by the node. Subsequently, based on the structural connection between the nodes, the matching degree between each pair of adjacent nodes is calculated. This matching degree reflects the similarity between the nodes in the feature space. After processing and normalization through the exponential function, the attention weight of each adjacent node to the current node is obtained. These weights determine which neighbors contribute more to the feature update of the current node. Finally, the feature information of each neighboring node is multiplied by the corresponding attention weight, and the results are summed up to obtain a new representation of the current node. This new representation integrates the feature information of the adjacent nodes in the structure and is weighted and adjusted according to the similarity, so that the current node can obtain richer local context information, thereby having a higher representation capability in the subsequent structural response prediction. The entire process is carried out node by node in the finite element mesh of the silicon steel sheet, retaining its topological connection relationship, and realizing the joint modeling of spatial, boundary and structural behavior characteristics.

[0107] S413: Based on the cross-matching feature representation vector, a local subgraph is constructed using the connection relationship between adjacent nodes in the graph, and the context features of each node are analyzed according to the weight value of the connected nodes to generate a unified fusion representation vector;

[0108] First, the output representation vector corresponding to each node is extracted and used as the local feature basis of the current node. Then, combined with the connection relationship in the silicon steel sheet graph structure, all adjacent nodes are searched for each node one by one to construct a local subgraph structure. The local subgraph consists of the central node and all its connected edges and adjacent nodes to form a local subgraph set. Then, the corresponding attention weight value is extracted for the node connection edge in each subgraph. The weight value comes from the weighted distribution in the cross-attention mechanism, which represents the contribution degree of each adjacent node to the central node feature. The context feature analysis operation is performed on each central node, that is, the cross feature vector is extracted from all its neighboring nodes, and is weighted multiplied with its weight value one by one. Then, a vector summation operation is performed, and the weighted sum of all weight values ​​multiplied by the corresponding neighbor feature vector is used as the context fusion representation of the central node. The above process is repeated until all nodes are locally fused, and finally a unified fusion representation vector set is generated. Each item represents the integrated feature state of a node in the silicon steel sheet structure considering the influence of its adjacent topology.

[0109] See also Figure 6 The specific steps for obtaining the performance prediction results of silicon steel sheets are as follows:

[0110] S511: Inputting the unified fusion representation vector into the multi-layer perceptron, the unified fusion representation vector corresponding to the graph embedding expression of each node in the silicon steel sheet grid topology structure, inputting the vector into the successive network layers of the perceptron in the order of node index, performing layer-by-layer vector mapping and nonlinear activation operations, and generating node stress prediction results;

[0111] In the finite element mesh modeling task of silicon steel sheets, the fusion representation vector of each node output by the cross-attention mechanism contains information such as its relative position in the topological structure, boundary response, and adjacent influence. This vector will be fed as input into the multi-layer perceptron (MLP) model for stress prediction tasks.

[0112] Set up the first The fusion representation vector of nodes is: ,in, : Node number, indicating its topological position in the silicon steel sheet grid; :node The fusion representation vector is input to the first layer of the perceptron; : The fusion vector dimension is determined by the graph structure output and contains spatial, boundary and structural response information.

[0113] Multilayer Perceptron Each layer performs a linear transformation and activation operation. The intermediate output of the layer is , then its dimension is: ;

[0114] Representation node In the The output representation of the hidden unit of the layer is recursively calculated as follows:

[0115] ;

[0116] ;

[0117] ;

[0118] .

[0119] The final output is a node The stress prediction value of: .

[0120] in, : No. The weight matrix of the layer; : No. The bias vector of the layer;

[0121] : Activation function (such as ReLU, Tanh), introducing nonlinear expression; : Input layer dimension, that is, fusion vector dimension; : Output layer dimension, indicating that the predicted stress is a scalar value.

[0122] Predicted value Corresponding silicon steel sheet node The stress response value under a given thermal cycle or load number, such as the magnitude of the principal stress or the equivalent stress in a specific direction. The predicted values ​​are numbered and organized into a vector for error calculation and reverse updating with the simulated stress values. This structure allows for multi-layered abstraction, gradually extracting the most relevant information mapping relationships from the node fusion features to the stress state. This makes it suitable for capturing the microscopic response patterns of silicon steel sheet structures under complex loading conditions.

[0123] In the thermal cycle and stress field simulation prediction task of silicon steel sheet structure, consider a certain node, and its fusion representation vector after graph network encoding is: ; indicates a node The encoding features of , where the first dimension represents the position features of the compressed projection of the node in the three-dimensional structure; the second dimension represents the embedded features of the boundary response and adjacency information.

[0124] This vector is input into a 2-layer MLP (containing 1 hidden layer and 1 output layer) for stress prediction.

[0125] First layer (input→hidden):

[0126] Weight Matrix : ;

[0127] Bias vector : ;

[0128] The activation function uses ReLU, that is, for each component ,Pick ;

[0129] Second layer (hidden → output):

[0130] Weight Matrix : ;

[0131] Bias term ;

[0132] The output dimension is 1, that is, the nodal stress prediction is a scalar.

[0133] First layer linear transformation + activation ;

[0134] Compute matrix multiplication: ;

[0135] Add the bias: ;

[0136] Activation (ReLU): ;

[0137] The second layer output calculation:

[0138] ;

[0139] The final predicted value is: ;

[0140] This value indicates that under the current thermal load number, the silicon steel sheet numbered The predicted grid node stress is -4 (the unit depends on the training sample simulation settings, such as MPa). Although this value is negative, its physical meaning, such as the direction of compressive stress, must be determined in conjunction with the actual stress sign definition in practical applications.

[0141] First, the fusion features obtained by graph network encoding of each silicon steel sheet node are used as the input vector, which represents the overall information of the node's position in the structure, boundary response and adjacent influence. The input vector first passes through the first layer of linear mapping and nonlinear activation operations to output the intermediate features of the hidden layer, representing the preliminary extracted local stress correlation features. Then, the hidden layer features continue to be passed to the next layer, repeating the same mapping and activation process, and continuously enhancing the nonlinear fitting ability between the input vector and the target stress value. In the final output layer, the last layer of hidden state is mapped to a scalar result through a linear transformation. This value is the stress prediction value of the current node, representing its local stress response intensity under specific loading conditions. The essence of the entire calculation process is to extract potential expressions that are highly correlated with the target stress from the initial fusion features through layer-by-layer weight transformation and activation function, so as to achieve accurate stress prediction modeling for each node in the silicon steel sheet structure.

[0142] S512: Based on the node stress prediction results, call the simulation information in the structured training sample set, match the corresponding target stress value according to the node index, establish a mapping relationship between the predicted value and the target value, calculate the residual term and loss value in the root mean square error function, and generate the silicon steel sheet performance prediction results;

[0143] According to the node stress prediction value output by the multi-layer perceptron, it is matched with the actual simulated stress value recorded in the structured training sample set, and the residual mapping relationship between the predicted value and the target value is constructed, and the prediction error and overall loss of each node are further calculated.

[0144] Assume that the first The predicted value of a node is , the true target stress value is , then the squared residual term of the node is: ;

[0145] Calculate the root mean square error function for all nodes, the expression is:

[0146] ;

[0147] in, :node The predicted stress value comes from the output of the multi-layer perceptron; :node The simulation target stress value comes from the structured training sample; :node The squared residual term measures the prediction error; : the total number of nodes in the silicon steel sheet mesh; RMSE: the square root of the global mean error, used to measure the overall fitting quality of the model.

[0148] If the node The predicted value is , while the actual simulation value is , then the residual is:

[0149] ;

[0150] Suppose there are two more nodes whose predictions and true values ​​are as follows:

[0151] Node 2: ;

[0152] Node 3: ;

[0153] Then the root mean square error of all nodes is:

[0154] ;

[0155] This value indicates that the model has a large deviation in this round of prediction, especially the serious prediction error from node 1, which forms a large single-point residual. The constructed sequence will be used as the performance prediction result of silicon steel sheet.

[0156] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for rapid performance prediction of thin-film components based on deep learning, characterized in that: The following steps are involved: S1: Collect simulation information of silicon steel sheets under various load and thermal cycle conditions to obtain a structured training sample set; S2: establishing a grid topology structure of the silicon steel sheet according to the silicon steel sheet simulation information in the structured training sample set, encoding the grid topology structure, and outputting the encoding input elements; S3: Inputting the coding input elements into a deep neural network, extracting input information of the geometric positions of silicon steel sheet nodes and the relationship between the boundaries from the silicon steel sheet grid topology structure, and outputting Trunk network coding features and Branch network coding features of the input information through network coding; S4: Fusing the Trunk network coding features and the Branch network coding features, enhancing local features through the self-attention mechanism, and obtaining a unified fusion representation vector; The steps for obtaining the unified fusion representation vector are specifically as follows: S411: Fusing the trunk network coding features and the branch network coding features, mapping the node geometric position features in the trunk network coding features to the boundary interaction features in the branch network coding features according to the node index, performing concatenation and combination on the two types of feature vectors to generate a node-level joint coding feature sequence; S412: Inputting the node-level joint encoding feature sequence into a cross-attention mechanism, using the position index of the node in the topological structure as a query item, constructing a mutual attention weight relationship between the joint encoding features, and generating a cross-matching feature representation vector; S413: Based on the cross-matching feature representation vector, a local subgraph is constructed using the connection relationship between adjacent nodes in the graph, and the context features of each node are analyzed according to the weight value of the connected nodes to generate a unified fusion representation vector; S5: inputting the unified fusion representation vector into a multi-layer perceptron to predict the stress of the silicon steel sheet, calling the simulation information in the structured training sample set to evaluate the prediction error, and obtaining the silicon steel sheet performance prediction result; The steps for obtaining the silicon steel sheet performance prediction results are specifically as follows: S511: Inputting the unified fusion representation vector into a multi-layer perceptron, wherein the unified fusion representation vector corresponds to a graph embedding expression of each node in the silicon steel sheet grid topology structure, and inputting the unified fusion representation vector into successive network layers of the perceptron in the order of node indexes, performing layer-by-layer vector mapping and nonlinear activation, and generating a node stress prediction result; S512: Based on the node stress prediction result, call the simulation information in the structured training sample set, match the corresponding target stress value according to the node index, establish a mapping relationship between the predicted value and the target value, calculate the residual term and loss value in the root mean square error function, and generate the silicon steel sheet performance prediction result.

2. The method for rapid performance prediction of thin-film components based on deep learning according to claim 1, characterized in that: The structured training sample set includes load type, thermal cycle conditions, and node response data. The encoding input elements are specifically grid connection relationships, topological structure labels, and geometric positioning parameters. The Trunk network encoding features are specifically node spatial distribution patterns and global structural features. The Branch network encoding features are specifically boundary adjacency relationship features and local geometric features. The unified fusion representation vector includes local enhancement features, multi-layer feature aggregation results, and attention distribution weights. The silicon steel sheet performance prediction results include stress numerical distribution, prediction deviation indicators, and performance evaluation outputs.

3. The method for rapid performance prediction of thin-film components based on deep learning according to claim 1, characterized in that: The steps for obtaining the structured training sample set are specifically as follows: S111: Collect simulation information of silicon steel sheets under various load and thermal cycle conditions, including node space coordinates, unit connection relationships, and boundary constraint parameters. The node space coordinates correspond to the grid node positions in the silicon steel sheet geometry, the unit connection relationships correspond to the combination methods in the silicon steel sheet geometry, and the boundary constraint parameters correspond to the force and restricted area settings in the silicon steel sheet loading condition properties. Extract simulation input structure information according to the loading number and generate a node configuration and constraint information set. S112: Based on the node configuration and constraint information set, extract the principal stress direction and principal stress tensor of the corresponding node in each thermal cycle stage, arrange the principal stress direction vectors of each stage according to the thermal loading sequence, calculate the angle change value of the direction data of the same node between consecutive stages, and record the direction change path; S113: Integrate the node space coordinates, unit connection relationship, principal stress tensor, direction change path, boundary constraint parameters and loading number corresponding to each node to generate a structured training sample set.

4. The method for rapid prediction of sheet component performance based on deep learning according to claim 3 is characterized in that: The steps for obtaining the coding input elements are specifically as follows: S211: obtaining all node numbers and corresponding coordinate positions according to the silicon steel sheet simulation information in the structured training sample set, identifying connection paths between nodes based on unit connection relationships, and constructing a silicon steel sheet grid topology structure; S212: Based on the silicon steel sheet grid topology, extract the spatial direction information of each node within the local topology range, calculate the direction vector and Euclidean length of the edge between adjacent nodes, obtain the point normal vector in units of nodes as the point feature, use the edge direction and edge length as the edge feature, call the fast point cloud normal vector estimation algorithm to correct the point normal vector, and generate a geometric structure feature set; S213: Integrate the geometric structure feature set into a model input format in the form of a graph structure expression, and output the encoded input elements.

5. The method for rapid prediction of thin-film component performance based on deep learning according to claim 4 is characterized in that: The steps for obtaining the Trunk network coding features and the Branch network coding features are specifically as follows: S311: Inputting the node spatial coordinates in the encoded input elements into the Trunk network of the deep neural network, where the node spatial coordinates correspond to the geometric positions of the nodes in the silicon steel sheet grid topology structure, arranging the node numbers in the order of the graph structure while keeping the connection relationship unchanged, and using the relative distribution relationship between the nodes to capture the arrangement characteristics in the structural topology, generating the Trunk network geometric layout features; S312: Inputting the boundary constraint parameters, point normal vectors, edge directions, and edge lengths in the encoded input elements into the Branch network of the deep neural network, wherein the boundary constraint parameters correspond to the boundary interaction conditions of the silicon steel sheet structure, the point normal vectors represent the node direction attributes, and the edge directions and edge lengths reflect the directionality of the structural connection, thereby generating the Branch network boundary response characteristics; S313: Arrange the branch network boundary interaction features and the trunk network geometric position features into a multi-dimensional coding vector corresponding to the structural position index to obtain the trunk network coding features and the branch network coding features.

Images