Extrusion stress prediction method based on point-field double-path fusion physical model

Abstract

The application provides a kind of extrusion stress prediction method based on point-field double path fusion physical AI model, by combining global point-to-point prediction (MLP path) and local overall field prediction (GCN path), realize the high-precision prediction of stress-strain relationship of multi-layer structure of mobile phone screen in the process of extrusion loading.The application effectively solves the problems of low calculation efficiency, modeling difficulty, poor adaptability of nonlinear materials and other problems existing in the traditional finite element method in processing ultra-thin screen multilayer composite material, the accuracy of stress propagation modeling is significantly improved by double path fusion strategy, the standardization input of different grid structures is realized by using geometry template mapping, combined with small sample learning enhancement and physical constraint embedding, the calculation efficiency and the generalization ability of the model are greatly improved while ensuring the physical consistency of the prediction results, providing an efficient and reliable simulation solution for screen extrusion failure analysis.

Description

Technical Field

[0001] This invention relates to computer-aided engineering simulation and artificial intelligence technology, belonging to the field of product structure reliability analysis and artificial intelligence modeling, specifically to a method for predicting extrusion stress based on a point-field dual-path fusion physical model. Background Technology

[0002] In the current industry context of the widespread emergence of smartphones and foldable / ultra-thin screen devices, the reliability of the mobile phone screen, as a core component for user interaction, has become a critical issue in screen design, and compression failure has become one of the main forms of screen damage. When a device is subjected to external pressure, the difference in modulus between the layers of materials causes stress redistribution, resulting in localized stress concentrations at interlayer interfaces and geometric abrupt change areas (such as holes and edges). This ultimately leads to irreversible damage such as glass breakage, OCA adhesive delamination, or flexible circuit breakage. Therefore, accurately predicting the screen's compression response is of great significance for improving product reliability and extending its service life.

[0003] Currently, the industry generally utilizes simulation methods based on Computer-Aided Engineering (CAE) (such as Finite Element Analysis (FEA)). However, when dealing with multilayer composite material structures and highly nonlinear materials (such as hyperelastic layers), this method still faces several bottlenecks in practical applications: First, mobile phone single-layer thickness is thin, and the number of data nodes is large and inconsistent; Second, mobile phone screens have complex multilayer structures with complex stress transmission paths between layers, making model convergence difficult, easily leading to solution failure, low computational efficiency, and high cost; Third, pure data-driven approaches ignore mechanical laws, resulting in weak model generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based simulation method for mobile phone screen extrusion, which aims to solve the problems of low computational efficiency, high cost, modeling difficulty, and poor adaptability to nonlinear materials in traditional finite element simulation for foldable screen design.

[0005] To address the aforementioned technical problems, this invention proposes a method for predicting extrusion stress based on a point-field dual-path fusion physical model. It rapidly predicts boundary effects through global point-to-point prediction (MLP path); captures long-range stress transmission through cross-layer attention in local global field prediction (GCN path); and then fuses features from geometric and physical encodings using a gated attention mechanism, ultimately achieving real-time, high-precision simulation of extrusion failure. Simultaneously, it also utilizes geometric template mapping and PINN embedding to achieve data augmentation and improve model generalization ability, respectively. The specific technical solution provided by this invention is as follows:

[0006] A method for predicting extrusive stress based on a point-field dual-path fusion physical model includes the following steps:

[0007] Step 1: Obtain sample data and perform geometric template mapping and structural consistency processing;

[0008] Step 2: Construct a dual-path prediction model that combines the global point-to-point prediction path with the local global field prediction path to achieve high-precision stress prediction for multi-layer screen structures. Its prediction output includes nodal displacement field and nodal stress field.

[0009] Step 3: Fusion of the three coded streams, including the geometric coded stream, the physical coded stream, and the temporal coded stream:

[0010] The geometric coding stream is based on a unified set of three-dimensional discrete nodes obtained by geometric template mapping. It combines the structural connection information between nodes determined by spatial proximity to construct the corresponding mesh topology representation and encodes the node boundary condition attributes and geometric features together.

[0011] The physical coding stream is embedded in the training of the dual-path prediction model by a physical constraint neural network. The physical constraint neural network embeds four types of physical residuals—energy density function, equilibrium equation, constitutive relation, and boundary condition—into the loss function, thereby enabling the prediction results to actively satisfy physical laws.

[0012] The temporal coding stream uses the node-level stress prediction results obtained at each time step in step 2 as temporal input to construct a stress time series representation of the node during the loading process, and uses a temporal coding network to model the above stress time series.

[0013] Step 4: The extracted features of the triple-encoded stream are fused using a gated attention mechanism. A Hausdorff distance constraint term is introduced into the loss function to penalize the spatial offset of the maximum stress point, thereby improving the structural consistency and stability of the stress distribution.

[0014] Furthermore, step 1 specifically includes the following sub-steps:

[0015] Step 1.1: Obtain input data, including: geometric structure data, material parameter data, loading and boundary condition data, finite element reference data, and material energy density data;

[0016] Step 1.2: Establish a neutral layer parameterized coordinate system to map the irregular 3D mesh to a unified reference 2D parameter domain, and use this coordinate system as a reference standard to reconstruct and align the original irregular mesh. This includes a structurally consistent mapping strategy or a physically consistent interpolation strategy.

[0017] The structure-consistent mapping strategy encodes the 3D model into a fixed-structure input feature tensor in space.

[0018] The physical uniformity interpolation strategy constructs a regular spatial sampling grid in the physical space, and then maps or interpolates the physical information of the 3D model into the regular grid;

[0019] Step 1.3: Sample Expansion: Collect the experimentally measured sample material parameters and represent each group of material samples as a parameter vector; combine all sample material parameters into a sample matrix and normalize it; use principal component analysis to decompose the normalized parameter matrix to obtain several principal component directions; perturb the central parameter in each principal component direction to obtain a new parameter vector generated after perturbation.

[0020] Furthermore, step 2 is detailed as follows:

[0021] Step 2.1: Point-to-point prediction path: The inconsistent 3D finite element meshes in different samples are uniformly mapped into a set of 3D discrete nodes with fixed spatial semantics. Using these discrete nodes as the smallest prediction unit, a convolutional neural network is employed as the convolutional coding network to extract the local spatial features of each node. The extracted spatial features, along with the node's geometric position information and corresponding material parameters, are input into a multilayer perceptron to achieve rapid point-to-point prediction of the stress response of each node. The output layer introduces a Hooke's law residual term.

[0022] ;

[0023] in Indicates the total number of nodes. This represents the stress value predicted by the multilayer perceptron at the i-th node. This represents the elastic modulus of the material corresponding to the i-th node. Poisson's ratio is represented by the ratio for different layers of material. , Indicates the strain components in the principal direction. The strain component in the direction corresponding to the i-th node It is the square of the L2 norm;

[0024] Step 2.2: Patch-level GCN field prediction path: A local structural field prediction model is established based on the Graph Convolutional Network (GCN). A patch-level graph structure is constructed, that is, the 3D model of a single sample is divided into multiple local blocks. Each local block constitutes a graph structure, where nodes represent spatial discrete points within the block, including the point's coordinates, the material layer to which it belongs, material parameters, local geometric features, and boundary attribute information; edges are used to describe the topological adjacency relationships between nodes; through a three-layer GCN network, short-range, medium-range, and long-range connectivity features are extracted respectively, and combined with a cross-layer attention mechanism, the output is the patch-level stress field and displacement field prediction results.

[0025] Furthermore, step 3 is detailed as follows:

[0026] Step 3.1: Geometric Encoding Flow: Based on the unified three-dimensional discrete node set obtained through geometric template mapping in Step 1, and combined with the structural connection information determined by the spatial proximity relationship between nodes, a corresponding mesh topology representation is constructed. The node boundary condition attributes and geometric features are then encoded to obtain a local structural feature vector. This geometric encoding flow is used to standardize and unify the geometric input features of the dual-path prediction model described in Step 2.

[0027] The information in the local structural feature vector also includes the curvature tensor, which is calculated by relating node coordinates to their local neighborhood structure. Then, the calculated curvature tensor is normalized and mapped to generate a spatial weight mask. Thus, geometric coding features are obtained. The specific expression is:

[0028] ;

[0029] in, It is Hadamaji. For the first Local structural feature vectors of each node;

[0030] Step 3.2: Physical Encoding Flow: The physical constraint neural network includes a multi-layer loss function constraint module and a physical embedding module, as detailed below:

[0031] The multi-layer loss function constraint module constructs an energy density function to achieve nonlinear response constraints on material stress:

[0032] ;

[0033] Where W represents the energy stored per unit volume of material during deformation. , C1 and C2 are the first and second strain tensor invariants, C1 and C2 are the material constants of the Yeoh model, and N' is the order of the Yeoh hyperelastic model.

[0034] The physical embedding module introduces four types of residual constraint terms using automatic differentiation techniques:

[0035] Data error terms:

[0036]

[0037] in u represents the predicted physical field variable. obs This is actual observation data;

[0038] Residual terms of the equilibrium equation:

[0039]

[0040] For divergence operators, This refers to the force acting on a unit volume of an object;

[0041] Constitutive relation residuals:

[0042]

[0043] in, and These are the nodal stress and nodal displacement output by the point-field dual-path prediction model, respectively. The strain tensor is calculated from the predicted displacement field using the strain-displacement geometry. This represents the mapping relationship between strain and stress calculated based on a known material constitutive model;

[0044] Boundary condition residuals:

[0045]

[0046] in, It is a known displacement value specified on the displacement boundary. It refers to the known surface force value specified on the force boundary. The boundary unit normal vector;

[0047] Introducing an adaptive dynamic weighting strategy:

[0048]

[0049] It is the variance of the m-th loss term during recent training, where m is the total number of loss terms. It is the variance of the k-th loss term itself, which is the weight to be calculated.

[0050] Therefore, the total residual function is:

[0051]

[0052] in, , The weights corresponding to the loss function are calculated using an adaptive dynamic weighting strategy.

[0053] Finally, the encoded material physical parameters are used to generate channel weight coefficients for different material layers. The characteristic influence of each material layer is calculated using the following formula:

[0054]

[0055] in, Indicates the first The weighting coefficients of each channel, For the first The elastic modulus of the layer material This represents the total number of material layers; thus, the physical coding feature is obtained. Expression:

[0056]

[0057] in, For the first The physical parameter vector corresponding to each material channel;

[0058] Step 3.3: Temporal Encoding Stream: Using the node-level stress prediction results obtained at each time step of the point-to-point prediction path and the patch-level field prediction path in Step 2 as temporal input, construct the stress time series representation of the node during the loading process; for the first... The stress evolution sequence of a spatial node during the extrusion loading process is represented as follows:

[0059]

[0060] in, Represents a node In the Predicted stress state at each loading time step This represents the total number of time steps in the entire loading process;

[0061] The stress time series above is modeled using a time-series coding network based on the Transformer structure. The time-series coding process can be expressed as follows:

[0062]

[0063] in, These are time-series coding features.

[0064] The beneficial effects of this invention are as follows:

[0065] The high number of stacked layers in mobile phone screens leads to complex stress transfer paths between layers. Traditional finite element methods struggle to accurately capture long-range stress propagation patterns across layers, and suffer from insufficient accuracy and high computational costs. This invention proposes a point-field dual-path fusion method: it rapidly predicts boundary effects through global point-to-point prediction (MLP path) and combines it with local global field prediction (GCN path) to capture long-range stress transfer through cross-layer attention. This method effectively solves the problem of difficult stress propagation modeling in multi-layered structures. Without increasing complexity, it significantly improves computational efficiency and simulation stability by fusing MLP and GCN algorithms.

[0066] Currently, ultra-thin screens are generally micrometer-thick, requiring extremely high network density in extrusion simulations. This results in simulation models with hundreds of thousands or even millions of network nodes, a massive number. Furthermore, differences in mesh size and topology among different samples lead to input alignment difficulties and inconsistent outputs, resulting in low computational efficiency and significant data redundancy. To address these issues, this invention introduces a geometric template mapping mechanism. By using geometric templates, 3D models of different sizes, shapes, and mesh distributions are uniformly encoded into fixed-structure inputs and outputs in space, achieving standardization and efficiency in the modeling process. This method significantly reduces the impact of node size differences on model predictions, reduces repetitive modeling workload, and improves the generalization ability and computational efficiency of AI models.

[0067] In high-precision finite element simulations, high-quality training samples are extremely scarce due to the need for fine mesh generation and the high computational resources required for each sample. Traditional AI methods are difficult to apply to this scenario due to their large data requirements. Therefore, this invention proposes a small-sample learning enhancement strategy. Based on existing samples, it generates tens of thousands of training samples through material parameter PCA perturbation rules and transfer learning mechanisms, effectively alleviating the problem of insufficient samples and improving the model's generalization ability and prediction accuracy.

[0068] Mobile phone screens consist of stacked materials such as cover glass, flexible substrate, OCA adhesive, and touch module. The interlayer properties differ significantly, exhibiting high material nonlinearity. Traditional neural networks rely on data fitting, making it difficult to guarantee mechanical conservation, constitutive consistency, and boundary condition satisfaction, and hindering generalization to the true physical response under new structures and operating conditions. To address these issues, this invention introduces PINN physical information neural network embedding constraints, incorporating physical equations and boundary condition constraints into the AI ​​prediction process to ensure results conform to physical laws. This method effectively improves convergence and prediction accuracy under nonlinear conditions, guarantees physical consistency, and achieves reliable modeling of the mechanical behavior of complex screen materials. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating the framework of the method of the present invention.

[0070] Figure 2 This is a flowchart of the structure consistency mapping strategy.

[0071] Figure 3 Flowchart of the physical consistency interpolation strategy.

[0072] Figure 4 The diagram shows the stress field solution; where (a) is the top view and (b) is the cross-sectional view.

[0073] Figure 5 This is a schematic diagram of the predicted simulation results. Detailed Implementation

[0074] To better understand the purpose, structure, and function of this invention, the invention will be described in further detail below with reference to the accompanying drawings.

[0075] A method for predicting extrusive stress based on a point-field dual-path fusion physical model, such as Figure 1 As shown, it includes the following steps:

[0076] Step 1: Obtain sample data and perform geometric template mapping and structural consistency processing.

[0077] Step 1.1: Obtaining Input Data. This invention innovatively establishes an information mapping mechanism for highly nonlinear materials, loading complex nonlinear constitutive relations onto the AI ​​model as energy density curves. This avoids brute-force solutions, achieves rapid fitting of material constitutive relations, and significantly reduces computational complexity. The input data includes:

[0078] 1) Geometric structure data: This includes information such as node coordinates, layer thickness distribution, and interface relationships of the multi-layered screen structure. Through geometric template mapping, irregular 3D meshes (mesh structures with different sizes, shapes, or topologies) are transformed into a unified set of discrete 3D nodes. During this process, the model's 3D mesh is normalized and standardized into a set of nodes, ensuring that the mesh structure across different samples can be processed uniformly.

[0079] 2) Material parameter data: including the elastic modulus, Poisson's ratio, Yeoh hyperelastic model parameters (C1, C2, C3) of each layer of material, and an extended set of material parameters generated by PCA perturbation;

[0080] 3) Loading and boundary condition data: including extrusion load path, displacement boundary constraints, contact area definition (core extrusion zone, OCA (optically transparent adhesive layer), CPI (colorless polyimide layer), free surface, volume domain), etc.

[0081] 4) Finite element reference data (for monitoring / calibration): including von Mises stress, displacement field, strain field distribution, and maximum stress values ​​at key points, etc.

[0082] 5) Material energy density data: Energy density-strain curves obtained from experiments or FEA are used to construct physical residual constraints through automatic differentiation.

[0083] Step 1.2: Preprocessing of the 3D model. This invention provides a simulation data preprocessing method based on dual-space mapping to solve the problem of inconsistent model input caused by differences in geometric scale, network density, and distribution among different samples. This is achieved by establishing a neutral layer parametric coordinate system. ,in These represent the normalized parameterized coordinates along the two principal directions of the neutral layer (typically corresponding to the X and Y directions of the screen plane). They are used to map irregular 3D meshes (i.e., meshes with different sizes, shapes, or topologies) to a unified reference 2D parameter domain. Using these coordinates as a reference standard, the original irregular mesh is reconstructed and aligned to achieve standardized input of multi-layer screen structure parameters. This mapping includes two core strategies:

[0084] Strategy 1: Structure Consistency Mapping Strategy. This strategy encodes 3D models of varying sizes, shapes, and mesh distributions in space into a unified input feature tensor with a fixed structure, while avoiding filling or sampling the entire model into a very dense, inefficient, large-volume point cloud / voxel mesh. The specific process is as follows: Figure 2 As shown, the 3D model is first normalized and centered to the cube frame. Then, the cube is divided into K³ local patches. Within each patch, a fixed-dimensional feature vector is extracted through interpolation sampling. Finally, the entire 3D model is encoded using a 3D CNN (3D Convolutional Neural Network), and all local features (i.e., the feature vectors of each patch) are concatenated into a unified input tensor, which is used as the input to the neural network.

[0085] Strategy Two: Physically Consistent Interpolation Strategy. A regular spatial sampling mesh is constructed within the physical space, and then the physical information of the 3D model is mapped or interpolated onto the regular mesh to achieve a unified input representation for different models. The specific process is as follows: Figure 3 As shown, the same process is performed first, with center alignment and normalization. Then, interpolation is performed from the centroid of the 3D model towards the edges, and interpolation is also performed on the boundary regions to construct a spatial network. Another approach is to adjust the resolution of the spatial network region through non-uniform sampling, and then interpolate and fill the sample points by combining the original 3D model network properties (including the geometric position, material properties, boundary conditions, and other physical characteristics of each node in the model), and finally output a tensor with a unified structure.

[0086] Step 1.3: Sample Expansion. Collect the experimentally measured material parameters of the samples, including the elastic modulus. Poisson's ratio and hyperelastic constitutive parameters (such as (etc.), and represent each group of material samples as a parameter vector:

[0087]

[0088] in, A vector representing all parameters of the i-th material sample; This indicates the elastic modulus of a material (Young's modulus). Indicates the Poisson's ratio of the material; The parameters represent those that describe the hyperelastic constitutive model of the material.

[0089] All sample material parameters were combined into a sample matrix and normalized to eliminate the influence of dimensional differences on the perturbation results. Principal component analysis (PCA) was then used to decompose the normalized parameter matrix to obtain several principal component directions. In each principal component direction Above, with central parameters Perturb the base point and obtain the new parameter vector generated after perturbation according to the following formula. :

[0090]

[0091] in, This is the disturbance coefficient, and its value range is set according to actual needs.

[0092] Using this method, the training samples are expanded by perturbing material parameters (±5%), time series scaling (±20%), and noise injection on the basis of the original limited metadata, thereby improving the model's generalization ability.

[0093] Step 2: Model architecture design; This invention introduces the "point-field dual-path fusion" strategy for the first time. By combining the global point-to-point prediction path with the local global field prediction path, high-precision stress prediction of multi-layer screen structures is achieved. Its prediction output includes nodal displacement field and nodal stress field, which are used to fully describe the mechanical response state of the screen under extrusion loading conditions, as follows.

[0094] Step 2.1: Point-to-point path prediction (MLP Path);

[0095] Based on the geometric template mapping and structural consistency processing completed in step 1, the inconsistent 3D finite element meshes from different samples are uniformly mapped into a set of 3D discrete nodes with fixed spatial semantics. Using these discrete nodes as the minimum prediction units, a convolutional neural network (CNN) is employed as the convolutional coding network to extract the local spatial features of each node. The extracted spatial features, along with the node's geometric location information and corresponding material parameters, are input into a multilayer perceptron (MLP) to achieve point-to-point rapid prediction of the stress response of each node. The output layer incorporates a Hooke's law residual term.

[0096]

[0097] in Indicates the total number of nodes. This represents the stress value predicted by the multilayer perceptron at the i-th node. This represents the elastic modulus of the material corresponding to the i-th node. Poisson's ratio varies depending on the material in each layer, and is expressed as... , Indicates the strain components in the principal direction. The strain component in the direction corresponding to the i-th node The square of the second norm is used to apply a squared penalty to the difference between the predicted stress and the theoretical stress.

[0098] This step still introduces the Hooke's Law residual term in the point-to-point prediction path as a weak physical prior constraint. It is used to stabilize the stress-strain mapping relationship in the early stage of model training and in the local linear elastic response range, so as to avoid point-level prediction from degenerating into pure data fitting, thereby improving the convergence stability and extrapolation generalization ability of the overall model.

[0099] The output of the multilayer perceptron is a node-level displacement vector and the corresponding stress components, enabling point-to-point prediction of the mechanical response of local nodes.

[0100] Step 2.2: Patch-level GCN field prediction path (GCN Path)

[0101] A local structural field prediction model is established based on a Graph Convolutional Network (GCN). A patch-level graph structure is constructed, dividing the 3D model of a single sample into multiple local patches. Each local patch constitutes a graph structure, where nodes represent spatial discrete points within the patch, including information such as coordinates, material layer, material parameters, local geometric features, and boundary attributes. Edges describe the topological adjacency relationships between nodes, such as connection methods, distance weights, and inter-layer transfer characteristics. Each local patch corresponds to a sub-region in the 3D model. The graph structure connects these patches through nodes and edges, forming a local graph that can be processed by a GCN. A three-layer GCN network extracts short-range (1-hop), mid-range (2-hop), and long-range (3-hop) connectivity features, and combines this with a cross-layer attention mechanism to further model stress transfer patterns, thereby improving the prediction accuracy of local high-stress regions.

[0102] The output of the graph convolutional network path is the Patch-level stress and displacement field prediction results, which are used to characterize the overall continuity of stress distribution and interlayer transfer characteristics in a local area.

[0103] Step 3: Triple Encoding Stream Fusion; Based on the point-field dual-path prediction model constructed in Step 2, in order to further unify the model input expression, embed physical consistency constraints, and characterize the load time series effect, this invention introduces geometric encoding stream, physical encoding stream, and time series encoding stream to collaboratively encode and fuse the input features, physical constraint information, and time series information of the dual-path prediction model, so as to improve the accuracy, stability, and generalization ability of the model in multi-layer screen extrusion stress prediction.

[0104] Step 3.1: Based on the unified 3D discrete node set obtained through geometric template mapping in Step 1, and combined with the structural connection information determined by the spatial proximity relationship between nodes, a corresponding mesh topology representation is constructed. This representation is then encoded using node boundary condition attributes and geometric features to obtain a local structural feature vector. This geometric encoding stream is used to standardize and unify the geometric input features of the dual-path prediction model described in Step 2, ensuring that although the MLP path and GCN path differ in input form, they can share the same geometric expression space at the geometric feature level. This unified geometric input representation ensures improved accuracy of local stress prediction while maintaining global geometric continuity.

[0105] The information in the local structural feature vector also includes the curvature tensor, which can be calculated using node coordinates and their local neighborhood structure. Specifically, the curvature tensor... In the The calculation at each node can be performed using the following formula:

[0106]

[0107] in, Let be the coordinates of the node. Let be the path length of the neighborhood structure. The curvature tensor of each node can be obtained by calculating the second derivative.

[0108] Then, after normalizing and mapping the calculated curvature tensor, a spatial weight mask can be generated. This mask is used to highlight areas of geometric abrupt changes and high stress risk. The normalization process can be performed using the following formula:

[0109]

[0110] in, The mask represents the maximum value of the curvature tensors of all nodes after normalization. This indicates the degree of stress risk in different areas.

[0111] Thus, geometric coding features are obtained. The specific expression is:

[0112]

[0113] in, It is Hadamaji. For the first The local structural feature vectors of each node extracted from the geometric coding stream in step 3.1 This is a spatial weight mask.

[0114] Step 3.2: Physical Encoding Flow: To ensure the physical consistency of the model output, this invention embeds a Physically Informed Neural Network (PINN) structure into the training of the dual-path prediction model. This structure consists of a multi-layer fully connected network, which obtains high-order gradients through automatic differentiation and embeds four types of physical residuals—energy density function, equilibrium equation, constitutive relation, and boundary conditions—into the loss function. This achieves proactive satisfaction of the prediction results with physical laws, fundamentally guaranteeing the physical consistency of the prediction. The Physically Informed Neural Network includes a multi-layer loss function constraint module and a physical embedding module, as detailed below:

[0115] The multi-layer loss function constraint module is based on the Yeoh hyperelastic model and achieves nonlinear response constraints on material stress by constructing an energy density function.

[0116]

[0117] Where W represents the energy stored per unit volume of material during deformation. , C1 and C2 are the first and second strain tensor invariants, respectively. C1 and C2 are the material constants of the Yeoh model, and N' is the order of the Yeoh hyperelastic model.

[0118] The order N' represents the complexity of the nonlinear stress-strain relationship described in the Yeoh hyperelastic model. The value of N' determines the degree of nonlinearity of the stress response of the material model when the strain is large. The higher the order, the more accurately the model can fit the complex stress-strain curve of the material.

[0119] The physical embedding module utilizes automatic differentiation technology to embed the following physical laws as loss functions during training, in order to better satisfy the mechanical conservation and constitutive relations in stress solving:

[0120]

[0121] For divergence operators, For Cauchy stress tensor, This represents the force acting on a unit volume of an object. For strain tensor.

[0122] This leads to the introduction of four types of residual constraint terms:

[0123] Data error terms:

[0124]

[0125] in u represents the predicted physical field variable. obs This is actual observation data.

[0126] Physical residual terms (three terms in total):

[0127] Residual terms of the equilibrium equation:

[0128]

[0129] Constitutive residual terms (nonlinear materials):

[0130]

[0131] in, and These are the nodal stress and nodal displacement output by the point-field dual-path prediction model, respectively. The strain tensor is calculated from the predicted displacement field using the strain-displacement geometry. This represents the mapping relationship between strain and stress calculated based on a known material constitutive model (such as the Yeoh hyperelastic model).

[0132] Boundary condition residuals (displacement / force boundary):

[0133]

[0134] in, It is a known displacement value specified on the displacement boundary. It refers to the known surface force value specified on the force boundary. This is the boundary unit normal vector.

[0135] Since the three types of physical residuals have different dimensions and convergence speeds in different samples and stages, this paper introduces an adaptive dynamic weighting strategy to improve the loss fusion effect. The weights corresponding to the three types of residual losses are as follows: Automatically adjusted by the following formula:

[0136]

[0137] It is the variance of the m-th loss term during recent training, where m is the total number of loss terms. It is the variance of the k-th loss term itself, which is the weight to be calculated.

[0138] This strategy can automatically balance gradient propagation based on the variance changes of each loss term, avoid a single physical constraint dominating the training process, improve the overall network convergence stability, and construct a stress prediction model with high robustness and high physical consistency.

[0139] Therefore, the total residual function is:

[0140]

[0141] in, , These are the weights corresponding to the loss function.

[0142] Finally, the encoded material physical parameters (such as elastic modulus) are... Poisson's ratio and hyperelastic material parameters (e.g., those of the Yeoh model). This is used to generate channel weight coefficients for different material layers. The characteristic influence of each material layer can be quantified by the material stiffness, specifically calculated using the following formula:

[0143]

[0144] in, Indicates the first The weighting coefficients of each channel, For the first The elastic modulus of the layer material This represents the total number of material layers. In this way, the channel weighting coefficients are dynamically adjusted based on the differences in material stiffness, thus obtaining the physical coding features. Expression:

[0145]

[0146] in, It is Hadamaji. Indicates the first The weighting coefficients of each channel, For the first The physical parameter vectors corresponding to each material channel (including elastic modulus, Poisson's ratio, and hyperelastic material parameters).

[0147] Step 3.3: Temporal Coding Flow: In view of the dynamic characteristics of load-stress response evolution over time during screen compression loading, this invention introduces temporal coding flow on the basis of geometric coding flow and physical coding flow to model the stress evolution law throughout the loading process, so as to characterize the time dependency relationship between different loading stages and improve the model's adaptability to nonlinear loading paths and loading rate changes.

[0148] Using the node-level stress prediction results obtained at each time step of the point-to-point prediction path (MLP Path) and patch-level field prediction path (GCN Path) in step 2 as time series inputs, a time series representation of the stress at the node during the loading process is constructed. For the first... The stress evolution sequence of a spatial node during the extrusion loading process is represented as follows:

[0149]

[0150] in, Represents a node In the Predicted stress state at each loading time step This represents the total number of time steps in the entire loading process.

[0151] A Transformer-based temporal coding network is used to model the stress sequence described above. This network employs a multi-head self-attention mechanism to weight the stress state at different time steps, thereby capturing the long-range temporal dependencies between different stages of the loading process and avoiding the gradient vanishing problem that traditional recurrent neural networks often encounter in long sequence modeling. Its temporal coding process can be represented as follows:

[0152] in, Temporal coding features: Through this temporal coding flow design, the model can accurately model the stress-strain relationship at different time points during loading. Especially under complex loading conditions, the model can better reflect the nonlinear response of the material, improve the accuracy of stress prediction, and adapt to the influence of different loading speeds, loading stages, and external conditions on material behavior.

[0153] Step 4: Feature Fusion and Model Optimization

[0154] To further improve the accuracy and physical consistency of the model, this invention proposes a stress prediction method based on feature adaptive weight fusion and spatial consistency optimization, on the basis of triple encoded stream output.

[0155] Step 4.1: Fusion Strategy:

[0156] After completing the feature extraction of the geometric coding stream, physical coding stream, and temporal coding stream, in order to fully integrate information from different dimensions and improve the prediction accuracy of the model under complex screen extrusion conditions, this invention constructs a feature fusion strategy based on spatial-channel-temporal adaptive weighting, which is used to comprehensively utilize geometric structural characteristics, material property differences, and temporal evolution information during the loading process.

[0157] The corresponding geometric coding features were obtained in step 3 above. Physical coding characteristics Time-series coding features Based on the three types of coding features, this invention employs a gated attention mechanism to adaptively weight and fuse geometric coding features, physical coding features, and temporal coding features. The fusion form is expressed as follows:

[0158]

[0159] in, The feature-gated fusion operator is used to adaptively adjust the contribution ratio of various features in stress prediction based on geometric complexity, material property differences, and temporal characteristics of the loading stage.

[0160] This fusion strategy can dynamically adjust the contribution ratio of each feature channel in stress prediction based on geometrical abrupt changes, differences in material properties, and temporal loading characteristics, thereby improving the accuracy and stability of high-stress region prediction.

[0161] Step 4.2: Optimization Strategy: Improving Accuracy through Key Point Coverage Optimization: Screen compression failure originates from stress singularities (crack tips / interlayer interfaces), and its spatial location error directly affects the reliability of failure prediction. A Hausdorff distance constraint term is introduced into the loss function to penalize the spatial offset of the maximum stress point (such as crack tips or recessed edges), thereby improving the structural consistency and stability of stress distribution.

[0162]

[0163] P is the set of predicted maximum stress points; Q is the set of actual failure initiation points; p and q correspond to a point in sets P and Q, respectively.

[0164] Enhancing generalization ability: By using a physical-guided generative adversarial network (GAN), perturbations at different levels are input, including material property perturbations, Young's modulus perturbations, and motion change perturbations, and new sample sets are obtained through physical verification. The final accuracy on the test set exceeds 85%.

[0165] Real-time computing acceleration: It adopts a combination of multi-dimensional hybrid parallel architecture and memory optimization strategy to achieve efficient resource utilization through distributed computing.

[0166] To verify the effectiveness of the method of the present invention, the present invention uses spherical extrusion as the classic loading method and uses ABAQUS software to construct a simulation model of the multi-layer structure of the screen based on finite element analysis (FEA).

[0167] The model uses C3D8RH eight-node three-dimensional solid elements, employs hybrid control to adapt to the properties of incompressible materials, enables reduced integrals to improve computational efficiency, and combines hourglass control to avoid numerical instability.

[0168] The mesh is structured, with the main screen (Part-1) containing 9375 elements and the loading component (Part-2) containing 1428 elements, to ensure sufficient mesh density in key stress areas, thereby improving simulation accuracy.

[0169] Regarding material properties: the screen OCA layer uses the Yeoh hyperelastic material model, while the remaining CPI layers use the linear elastic model;

[0170] Regarding loading and boundary condition settings: For stacked module materials, the bottom four boundary lines are subjected to completely fixed constraints, the ball is subjected to displacement in the z direction, and other degrees of freedom are restricted to ensure that the boundary conditions conform to the actual assembly state.

[0171] Finite element simulation results: such as Figure 4 The stress field distribution characteristics shown are as follows: Figure 4 (a) is a top view of the 15 sets of solution results. The extrusion stress is transmitted in a concentric circle, with the maximum stress concentrated at the extrusion center (yellow hot zone). The critical yield strength decreases outward in a concentric circle (blue-green to dark blue area), indicating that the OCA adhesive layer dissipates the extrusion impact energy in a stepwise manner. Furthermore, the stress cloud diagram shows that the stress distribution is symmetrical along the X / Y axes with a deviation of less than 3%, consistent with the theoretical assumptions of the screen neutral layer. The smooth stress field gradient (without abrupt changes) also proves that the OCA adhesive effectively transmits shear strain and avoids stress concentration in the brittle layer.

[0172] Figure 4 The stress profile in (b) shows a five-layer CPI / OCA overlapping structure (each three columns represent one material): the high stress area is concentrated in the rigid CPI film layer, and the OCA adhesive layer (blue-green) bears the shear deformation.

[0173] Forty sets of multilayer extrusion response data were generated using finite element simulation. Of these, 30 sets were used for training the method, and 10 sets were used to verify the model's generalization ability. All input data included geometric information (node ​​coordinates, thickness distribution), material constitutive parameters (Young's modulus, hyperelastic parameters C1, C2, C3), and loading conditions. The mesh and output stress distribution were transformed into tensor formats that could be directly input into the neural network for AI simulation through geometric template mapping and normalization.

[0174] Figure 5 The middle section presents nine examples of prediction results from the method of this invention, which are compared with... Figure 4 The comparison shows that the stress cloud distribution of the simulation results obtained by this method is completely consistent with the color distribution of the results obtained by traditional finite element simulation, proving that this model strictly follows the physical laws of extrusion simulation and the stress attenuation mode is highly consistent with the finite element solution. At the same time, the stress distribution of the simulation results obtained by this method is smoother. Although it loses some local accuracy to a certain extent, it avoids the stress oscillation caused by mesh distortion in traditional FEA.

[0175] Table 1: Verification of the prediction accuracy of AI simulation and finite element analysis.

[0176] Group number Maximum stress value (MPa) using this method FEA (Maximum Stress Value) (MPa) Total error error 2 2.17 2.502 -0.332 -13.3% 7 2.16 1.712 0.448 26.2% 14 2.16 1.993 0.167 8.4% 20 2.17 1.953 0.217 11.1% 30 2.16 1.997 0.163 8.2% 33 2.22 2.934 -0.714 -24.3% Mean \ \ -0.0085 10.82%

[0177] Table 1 selects six representative sets of data to compare the maximum stress values ​​of AI simulation and finite element simulation under the same conditions. It can be seen that AI simulation has a certain stability advantage and a smaller predicted fluctuation range, while FEA shows significant fluctuations. The AI ​​model exhibits an underestimation effect in the high stress zone (group number 33, with an error of 24.3%), resulting in a significant increase in systematic error. However, in other stress ranges, AI simulation has high accuracy.

[0178] However, given the same workload, the AI ​​simulation method proposed in this invention has a significant computational advantage. The average time taken by traditional FEA is hundreds of times longer than that of this method. Furthermore, according to the comparison in Table 2, the method proposed in this invention also has advantages compared to other AI models: MLP has low computational cost but high error, while CNN+RNN has long computation time. This invention combines the advantages of both, achieving a dual consideration of computational cost and error.

[0179] Table 2 Comparison of Model Time Consumption

[0180] Model Feature fusion method Average computation time (s) error(%) This method Space-Channel Dual-Gate Control 58 9.2 Pure MLP Fully connected layer 22 17.5 CNN+RNN Temporal convolution 143 12.3

[0181] Case results show that the method of the present invention maintains high prediction accuracy while improving computational efficiency by tens of times compared with finite element simulation; the model can accurately characterize the influence of thickness and hyperelastic parameters on the distribution of extrusion stress; under different screen structures and parameter conditions, the present invention has good generalization ability and engineering application value.

[0182] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for predicting extrusive stress based on a point-field dual-path fusion physical model, characterized in that, Includes the following steps: Step 1: Obtain sample data and perform geometric template mapping and structural consistency processing; Step 2: Construct a dual-path prediction model that combines the global point-to-point prediction path with the local global field prediction path to achieve high-precision stress prediction for multi-layer screen structures. Its prediction output includes nodal displacement field and nodal stress field. Step 3: Fusion of the three coded streams, including the geometric coded stream, the physical coded stream, and the temporal coded stream: The geometric coding stream is based on a unified set of three-dimensional discrete nodes obtained by geometric template mapping. It combines the structural connection information between nodes determined by spatial proximity to construct the corresponding mesh topology representation and encodes the node boundary condition attributes and geometric features together. The physical coding stream is embedded in the training of the dual-path prediction model by a physical constraint neural network. The physical constraint neural network embeds four types of physical residuals—energy density function, equilibrium equation, constitutive relation, and boundary condition—into the loss function, thereby enabling the prediction results to actively satisfy physical laws. The temporal coding stream uses the node-level stress prediction results obtained at each time step in step 2 as temporal input to construct a stress time series representation of the node during the loading process, and uses a temporal coding network to model the above stress time series. Step 4: The extracted features of the triple-encoded stream are fused using a gated attention mechanism. A Hausdorff distance constraint term is introduced into the loss function to penalize the spatial offset of the maximum stress point, thereby improving the structural consistency and stability of the stress distribution.

2. The method for predicting extrusive stress based on a point-field dual-path fusion physical model according to claim 1, characterized in that, Step 1 specifically includes the following sub-steps: Step 1.1: Obtain input data, including: geometric structure data, material parameter data, loading and boundary condition data, finite element reference data, and material energy density data; Step 1.2: Establish a neutral layer parameterized coordinate system to map the irregular 3D mesh to a unified reference 2D parameter domain, and use this coordinate system as a reference standard to reconstruct and align the original irregular mesh. This includes a structurally consistent mapping strategy or a physically consistent interpolation strategy. The structure-consistent mapping strategy encodes the 3D model into a fixed-structure input feature tensor in space. The physical uniformity interpolation strategy constructs a regular spatial sampling grid in the physical space, and then maps or interpolates the physical information of the 3D model into the regular grid; Step 1.3: Sample Expansion: Collect the experimentally measured sample material parameters and represent each group of material samples as a parameter vector; combine all sample material parameters into a sample matrix and normalize it; use principal component analysis to decompose the normalized parameter matrix to obtain several principal component directions; perturb the central parameter in each principal component direction to obtain a new parameter vector generated after perturbation.

3. The method for predicting extrusive stress based on a point-field dual-path fusion physical model according to claim 2, characterized in that, Step 2 is described in detail below: Step 2.1: Point-to-point prediction path: The inconsistent 3D finite element meshes in different samples are uniformly mapped into a set of 3D discrete nodes with fixed spatial semantics. The discrete nodes are used as the smallest prediction units, and a convolutional neural network is used as the convolutional coding network to extract the local spatial features of each node. The extracted spatial features, along with the geometric position information of the nodes and the corresponding material parameters, are input into a multilayer perceptron to achieve point-to-point rapid prediction of the stress response of each node. The output layer introduces Hooke's Law residual terms: ； in Indicates the total number of nodes. This represents the stress value predicted by the multilayer perceptron at the i-th node. This represents the elastic modulus of the material corresponding to the i-th node. Poisson's ratio is represented by the ratio for different layers of material. , Indicates the strain components in the principal direction. The strain component in the direction corresponding to the i-th node It is the square of the L2 norm; Step 2.2: Patch-level GCN field prediction path: A local structural field prediction model is established based on the Graph Convolutional Network (GCN). A patch-level graph structure is constructed, that is, the 3D model of a single sample is divided into multiple local blocks. Each local block constitutes a graph structure, where nodes represent spatial discrete points within the block, including the point's coordinates, the material layer to which it belongs, material parameters, local geometric features, and boundary attribute information; edges are used to describe the topological adjacency relationships between nodes; through a three-layer GCN network, short-range, medium-range, and long-range connectivity features are extracted respectively, and combined with a cross-layer attention mechanism, the output is the patch-level stress field and displacement field prediction results.

4. The method for predicting extrusive stress based on a point-field dual-path fusion physical model according to claim 3, characterized in that, Step 3 is described in detail below: Step 3.1: Geometric Encoding Flow: Based on the unified three-dimensional discrete node set obtained through geometric template mapping in Step 1, and combined with the structural connection information determined by the spatial proximity relationship between nodes, a corresponding mesh topology representation is constructed. The node boundary condition attributes and geometric features are then encoded to obtain a local structural feature vector. This geometric encoding flow is used to standardize and unify the geometric input features of the dual-path prediction model described in Step 2. The information in the local structural feature vector also includes the curvature tensor, which is calculated by relating node coordinates to their local neighborhood structure. Then, the calculated curvature tensor is normalized and mapped to generate a spatial weight mask. Thus, geometric coding features are obtained. The specific expression is: ； in, It is Hadamaji. For the first Local structural feature vectors of each node; Step 3.2: Physical Encoding Flow: The physical constraint neural network includes a multi-layer loss function constraint module and a physical embedding module, as detailed below: The multi-layer loss function constraint module constructs an energy density function to achieve nonlinear response constraints on material stress: ； Where W represents the energy stored per unit volume of material during deformation. , C1 and C2 are the first and second strain tensor invariants, C1 and C2 are the material constants of the Yeoh model, and N' is the order of the Yeoh hyperelastic model. The physical embedding module introduces four types of residual constraint terms using automatic differentiation techniques: Data error terms: in u represents the predicted physical field variable. obs This is actual observation data; Residual terms of the equilibrium equation: For divergence operators, This refers to the force acting on a unit volume of an object; Constitutive relation residuals: in, and These are the nodal stress and nodal displacement output by the point-field dual-path prediction model, respectively. The strain tensor is calculated from the predicted displacement field using the strain-displacement geometry. This represents the mapping relationship between strain and stress calculated based on a known material constitutive model; Boundary condition residuals: in, It is a known displacement value specified on the displacement boundary. It refers to the known surface force value specified on the force boundary. The boundary unit normal vector; Introducing an adaptive dynamic weighting strategy: It is the variance of the m-th loss term during recent training, where m is the total number of loss terms. It is the variance of the k-th loss term itself, which is the weight to be calculated. Therefore, the total residual function is: in, , The weights corresponding to the loss function are calculated using an adaptive dynamic weighting strategy. Finally, the encoded material physical parameters are used to generate channel weight coefficients for different material layers. The characteristic influence of each material layer is calculated using the following formula: in, Indicates the first The weighting coefficients of each channel, For the first The elastic modulus of the layer material This represents the total number of material layers; thus, the physical coding feature is obtained. Expression: in, For the first The physical parameter vector corresponding to each material channel; Step 3.3: Temporal Encoding Stream: Using the node-level stress prediction results obtained at each time step of the point-to-point prediction path and the patch-level field prediction path in Step 2 as temporal input, construct the stress time series representation of the node during the loading process; for the first... The stress evolution sequence of a spatial node during the extrusion loading process is represented as follows: in, Represents a node In the Predicted stress state at each loading time step This represents the total number of time steps in the entire loading process; The stress time series above is modeled using a time-series coding network based on the Transformer structure. The time-series coding process can be expressed as follows: in, These are time-series coding features.

5. The method for predicting extrusive stress based on a point-field dual-path fusion physical model according to claim 4, characterized in that, The Hausdorff distance constraint term is as follows: P is the set of predicted maximum stress points, and Q is the set of actual failure initiation points. p and q correspond to a point in the sets P and Q, respectively.

6. The method for predicting extrusion stress based on a point-field dual-path fusion physical model according to claim 5, characterized in that, The specific structural uniformity mapping strategy and physical uniformity interpolation strategy are as follows: Structure-consistent mapping strategy: First, normalize the 3D model and center it to the cube frame. Then, divide the cube into K³ local blocks. In each block, extract fixed-dimensional feature vectors through interpolation sampling. Finally, use a 3D convolutional neural network to encode the entire 3D model and concatenate all local features into a unified input tensor. Physically consistent interpolation strategy: First, perform center alignment normalization processing, then interpolate from the centroid of the 3D model to the edge, and interpolate the boundary region to construct a spatial network. Another path is to adjust the resolution of the spatial network region through non-uniform sampling, and fill the sample points by interpolating the network properties of the original 3D model. The final output is a tensor with a unified structure.

7. The method for predicting extrusive stress based on a point-field dual-path fusion physical model according to claim 6, characterized in that, The geometric structure data includes node coordinates, layer thickness distribution, and interface relationships of the multi-layer screen structure; the material parameter data includes the elastic modulus, Poisson's ratio, and Yeoh hyperelastic model parameters of each layer material; the loading and boundary condition data includes the extrusion load path, displacement boundary constraints, and contact area definition; the finite element reference data includes von Mises stress, displacement field, strain field distribution, and maximum stress value at key points; and the material energy density data includes energy density-strain curves obtained from experiments or FEA.

Images