A general AI-CAE multiphysics simulation system based on SDF-Graph-Grid three-layer fusion

The AI-CAE multiphysics simulation system, which integrates SDF, Graph, and Grid, solves the adaptation problem of unstructured grids in existing technologies, realizes a unified architecture for multiphysics, supports complex CAD geometry and cross-media simulation, improves simulation accuracy and efficiency, and has the ability to predict new materials.

CN123116880APending Publication Date: 2026-11-17XIAN ZHONGLANG AL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611612390.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively adapt to unstructured meshes in industrial scenarios, cannot capture long-range physical coupling, lack a unified multiphysics architecture, cannot handle transient problems, have limited generalization capabilities, and physical consistency depends on data fitting.

Method used

A general AI-CAE multiphysics simulation system that integrates SDF, Graph, and Grid is adopted. It adapts to arbitrarily complex CAD geometry through CAD preprocessing, SDF encoding, graph construction, and inference engine. It supports multi-media processing, combines graph neural operators and frequency domain neural operators for global calculation, supports transient simulation, and ensures physical consistency through cross-domain edge refraction conditions.

Benefits of technology

It achieves native adaptation to arbitrarily complex CAD geometry, supports fluid, conjugate heat transfer, high-frequency electromagnetic and fluid-structure interaction simulations, has cross-material generalization capabilities, reduces hardware resource consumption, improves simulation efficiency and accuracy, satisfies physical causality laws, and supports arbitrary boundary conditions and new material predictions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116880A_ABST
    Figure CN123116880A_ABST
Patent Text Reader

Abstract

This invention presents a general-purpose AI-CAE multiphysics simulation system based on a three-layer fusion of SDF, Graph, and Grid, comprising CAD preprocessing, SDF encoding, graph construction, inference engine, and storage / output modules. The inference engine includes a three-model submodule and a mode selection unit. This invention utilizes SDF implicit geometric encoding and Graph topology to natively adapt to arbitrarily complex CAD geometry and unstructured meshes. It employs a segmented multi-medium SDF vector field to accurately distinguish multi-medium partitions and extract geometric features. The mode selection unit adaptively switches between the three models based on the simulation type and SDF gradient magnitude, balancing local fine-grained prediction with long-range global coupling. A unified graph propagation framework in the time dimension supports transient simulation of moving boundaries. Endogenous cross-domain edge refraction conditions ensure flux continuity, achieving efficient and high-precision simulation of complex geometric multiphysics fields for industrial applications.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of rapid multiphysics simulation technology that integrates neural operator methods and geometric topology modeling, and particularly relates to a general AI-CAE multiphysics simulation system based on the fusion of three layers: SDF-Graph-Grid. Background Technology

[0002] Traditional CAE simulation methods, including commercial software such as Fluent, Ansys, and COMSOL, follow a core solution process: users manually divide the computational grid based on the CAD geometric model, discretize the partial differential equations at the grid nodes, and solve large-scale linear equation systems through iterative methods (such as the SIMPLE algorithm, conjugate gradient method, and multigrid method) until convergence. This process has been used in industry for decades and has mature mathematical theory and software implementation. In recent years, rapid simulation methods based on artificial intelligence have gradually emerged, mainly divided into three technical routes. The first category is pure FNO (Fourier Neural Operator) or Grid-type methods, which represent the simulation domain as a regular structured grid and learn the mapping from input parameters to output physical fields through frequency domain convolution operations. The second category is pure Graph or GNO (Graph Neural Operator)-type methods, which use unstructured grid cells as graph nodes and learn the relationship between local neighborhood features and physical fields through graph convolutional networks. The third category is the GINO (Graph-Grid Neural Operator) academic model, which proposes a preliminary algorithmic framework combining Graph and Grid in published literature. In addition, SDF (symbolic distance field) has been applied in the field of computer graphics as a geometric representation method, and there are also scattered public studies on graph neural networks and FNO in their respective fields.

[0003] The aforementioned existing technologies each have inherent limitations, restricting their practical application in industrial-grade multiphysics simulations. First, pure FNO or Grid-based methods only support regular structured meshes, failing to adapt to the widely used unstructured CAE meshes in industrial scenarios. Furthermore, when encountering multi-media partitions, curved boundaries, or thin-layer structures, forcibly mapping unstructured meshes to regular meshes introduces significant numerical errors due to geometric distortion and material interface blurring. Additionally, mesh homogenization preprocessing itself requires substantial additional interpolation calculations and storage overhead. Second, while pure Graph or GNO-based methods can adapt to unstructured meshes, their message passing mechanisms operate only within local neighborhoods, making it difficult to capture macroscopic physical effects such as long-range wave propagation, electromagnetic resonance, and global convection heat transfer in fluids, resulting in insufficient prediction accuracy for large-scale simulations. Third, existing GINO academic models only disclose the basic algorithm network structure, lacking a complete industrial software preprocessing workflow (such as multi-part CAD repair and watershed extraction), multi-media material binding mechanisms, multiple physics adaptation interfaces, and a closed-loop system for result post-processing and visualization, making direct deployment in industrial design processes impossible. Fourth, in existing technologies, SDF geometric coding, graph neural networks, and FNO exist as isolated modules, lacking a unified geometry-topology-global coupling coding system. This results in a single model being unable to simultaneously meet the simulation requirements of different physical fields such as fluid dynamics, conjugate heat transfer, high-frequency electromagnetics, and fluid-structure interaction. Fifth, when dealing with transient problems involving geometric boundary movement (such as structural deformation and shock wave propagation in fluid-structure interaction), existing methods cannot achieve continuous adaptive adjustment of edge weights with geometric evolution while maintaining graph topological stability. They also lack temporal causal constraints to guarantee the physical causality of information propagation. Sixth, the material parameters of existing methods are input in discrete form, making it impossible to effectively predict new materials not included in the training set. Their generalization ability is limited, and their physical consistency relies solely on data fitting or an external penalty loss term, failing to intrinsically guarantee the continuity of interface flux from the propagation rule level. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of existing technologies, this invention provides a general AI-CAE multiphysics simulation system based on the fusion of SDF, Graph, and Grid. This system solves the problems of existing technologies, such as the lack of long-range coupling in pure FNO restricted regular grids, the lack of fusion schemes among SDF, Graph, and Grid, the lack of industrial closed loops, and the lack of a unified architecture for multiphysics.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A general AI-CAE multiphysics simulation system based on SDF-Graph-Grid three-layer fusion includes:

[0007] The CAD preprocessing module is used to perform mesh repair, watershed extraction and component partitioning on multi-component CAD models, and bind each component or each medium domain to the corresponding material physical property parameters;

[0008] The SDF encoding module is used to generate a segmented multi-media symbolic distance field and extract the geometric features of each node from the symbolic distance field.

[0009] The graph construction module is used to construct edges based on the adjacency relationships between discrete nodes, forming a graph structure. It also constructs cross-domain edges for nodes belonging to different medium domains but sharing an interface. The module also integrates the geometric features, material physical property parameters, and boundary condition labels into the initial feature vector of each node.

[0010] The inference engine includes a first inference submodule, a second inference submodule, a third inference submodule, and a mode selection unit;

[0011] The first inference submodule is used to take the graph structure and the initial feature vector as input, aggregate the neighborhood features of the nodes through graph neural operators, and output the physical field prediction value of each node.

[0012] The second inference submodule is used to take the graph structure and the initial feature vector as input, project the graph node features onto the implicit regular grid through the graph encoder, perform global calculation on the implicit regular grid through the frequency domain neural operator, and then project the calculation result back to the original graph node through the graph decoder to output the physical field prediction value of each node.

[0013] The third inference submodule is used to expand the three-dimensional spatial nodes into four-dimensional spatiotemporal nodes on the time axis, construct a four-dimensional spatiotemporal graph containing spatial edges and temporal edges, calculate the message passing weight of each edge according to the gradient of the symbolic distance field and perform iterative propagation, and output the physical field prediction value of the four-dimensional spatiotemporal node.

[0014] The mode selection unit is used to select the first inference submodule, the second inference submodule, or the third inference submodule to perform inference based on the simulation type and geometric features.

[0015] The storage and output module is used to write inference results, symbolic distance field features, graph topology, and material parameters into a data file.

[0016] Preferably, the selection rule of the mode selection unit is as follows: when the simulation type is steady state and there is a node in the graph structure whose sign distance field gradient magnitude is greater than three times the average gradient magnitude of all nodes in the region where the node is located, the first inference submodule is selected; when the simulation type is steady state and there is no such node, the second inference submodule is selected; when the simulation type is transient, the third inference submodule is selected.

[0017] Preferably, the geometric features extracted by the SDF encoding module include inner and outer domain masks, distances to walls, distances to medium interfaces, and boundary layer thickness.

[0018] Preferably, the graph neural operator is a graph neural operator network, which aggregates the neighborhood features of each node by multiplying the feature vector of each neighboring node of the node with a learnable weight matrix, and then multiplying by a radial basis function value. The independent variable of the radial basis function is the Euclidean distance between the node and its neighboring nodes. After summing all the products, the updated features of the node are output through a nonlinear activation function.

[0019] Preferably, the graph encoder is a graph attention network encoder or a graph convolutional network encoder, used to project graph node features onto an implicit regular grid; the graph decoder is a graph attention network decoder or a graph convolutional network decoder, used to back-project the computation results on the implicit regular grid back to the original graph nodes.

[0020] Preferably, the frequency domain neural operator is a Fourier neural operator, and its convolutional layer performs the following operations: transforming the field data on the implicit regular grid to the frequency domain via Fourier transform, multiplying it in the frequency domain with a learnable frequency domain weight matrix, and then transforming it back to the spatial domain via inverse Fourier transform.

[0021] Preferably, the third inference submodule performs four-dimensional spacetime graph propagation inference in the following specific manner:

[0022] Expand the three-dimensional spatial nodes on the time axis to generate a four-dimensional spatiotemporal node set, and construct spatial edges and temporal edges on this set;

[0023] Calculate the symbolic distance field value at each four-dimensional node and its four-dimensional gradient with respect to spatial and temporal coordinates. The geodesic tensor is formed by the outer product of these four-dimensional gradients.

[0024] For each edge, the message passing weight is calculated based on the quadratic value of the edge direction vector and the geodesic tensor. The weight is the value of a natural exponential function with the negative value of the product of the quadratic value and the preset scale parameter as the exponent. For temporal edges, if the time interval of the edge is less than the spatial distance divided by the maximum propagation speed of the physical field, the weight is set to zero.

[0025] The message passing iteration is performed with the weights, and the new message of each node is the sum of the old messages of the neighboring nodes multiplied by the corresponding edge weight and the geodesic curvature attenuation factor, wherein the geodesic curvature attenuation factor is the value of the natural exponential function with the negative value of the integral of the sign distance field isosurface curvature line along the edge path as the exponent.

[0026] For cross-domain edges, the line integral adopts the refraction condition at the medium interface, that is, the ratio of the sine values ​​of the refractive indices of the two media is equal to the ratio of the sine values ​​of the incident angle and the sine value of the refraction angle, and the refractive index is obtained by mapping the material physical property parameters of the medium.

[0027] Repeat the iteration until the maximum value of the relative change of messages of all nodes is less than the preset convergence threshold, and output the physical field prediction values ​​of all four-dimensional spatiotemporal nodes.

[0028] Preferably, the preset scale parameter ranges from 0.01 to 10.0, and the preset convergence threshold ranges from 0.0001 to 0.01. In the third inference submodule, the time step is adjusted as follows: in each time step, the absolute value of the time partial derivative of the symbol distance field of each node is calculated, and the maximum value is taken. If the maximum value is greater than the first threshold, the current time step is reduced; if the maximum value is less than the second threshold, the current time step is expanded. The first threshold is 0.1, and the second threshold is 0.01.

[0029] Preferably, the boundary condition labels include at least one of the following types: velocity inlet boundary, pressure outlet boundary, no-slip wall boundary, symmetric boundary, periodic boundary, heat flux boundary, convective heat transfer boundary, radiation boundary, lumped port excitation boundary, and wave port excitation boundary; the material physical property parameters include: fluid density, dynamic viscosity, and thermal conductivity; solid thermal conductivity, elastic modulus, and Poisson's ratio; electromagnetic medium relative permittivity, loss tangent, and wavenumber; and PML absorber layer electrical conductivity and magnetic permeability.

[0030] Preferably, the multiphysics simulation is one of the following types: fluid simulation, whose governing equation is the Navier-Stokes equation, and the viscosity and thermal conductivity in the equation are temperature-dependent parameters; conjugate heat transfer simulation, whose governing equation is a coupled equation of the Navier-Stokes equation and the energy equation, and the thermal conductivity of the solid is a temperature-dependent parameter; high-frequency electromagnetic simulation, whose governing equation is the Maxwell equation, and the dielectric constant and permeability are frequency-dependent parameters; fluid-structure interaction simulation, whose governing equation is a coupled equation of the Navier-Stokes equation and the solid mechanics equation, and the elastic modulus of the solid is a temperature-dependent parameter.

[0031] The technical effects and advantages of the general AI-CAE multiphysics simulation system based on the SDF-Graph-Grid three-layer fusion of this invention are as follows:

[0032] 1. Architectural versatility and industrial applicability: This invention, through a three-layer fusion architecture of SDF implicit geometric encoding, Graph topology construction, and Grid frequency domain inference, achieves for the first time native adaptation to arbitrarily complex CAD geometries (including multi-media partitions, curved surface boundaries, and thin-layer structures) without the need for regular mesh homogenization preprocessing. This architecture simultaneously solves the problems of pure FNO methods being unable to adapt to unstructured meshes and pure GNO methods being unable to capture long-range physical couplings. A single model uniformly supports four industrial scenarios: fluid simulation, conjugate heat transfer simulation, high-frequency electromagnetic simulation, and fluid-structure interaction simulation.

[0033] 2. Accuracy in multi-media processing: This invention unifies the material physical property parameters of fluid media, solid structures, electromagnetic media, and PML absorbing layers as input feature channels for graph nodes, enabling the model to generalize across materials and media. The introduction of a piecewise multi-media SDF vector field allows for the accurate expression of the interface positions and topological relationships between different media domains, avoiding numerical diffusion and interface ambiguity introduced by mask approximation.

[0034] 3. Balancing inference accuracy and efficiency: This invention uses a mode selection unit to automatically switch inference modes based on the SDF gradient magnitude. In high gradient regions (boundary layers, interfaces), pure Graph local fine-grained inference is enabled to ensure local accuracy; in gradually varying regions, Graph→Grid mapping global inference is enabled to ensure consistency of long-range physical coupling. This adaptive scheduling of the two modes enables the model to simultaneously possess both local fine-grained resolution and global physical field consistency.

[0035] 4. Native support for transient moving boundaries: This invention unifies the time and space dimensions into a four-dimensional spatiotemporal graph. Using the covariant gradient tensor of the SDF spatiotemporal field as the routing driver for graph message passing, edge weights continuously evolve with the movement of geometric boundaries, eliminating the need for discrete reconnection of the graph topology. Causal cone constraints on temporal edges ensure that information propagation satisfies physical causality, maintaining the model's physical reliability in transient scenarios with moving boundaries, such as fluid-structure interaction and moving shock waves.

[0036] 5. Intrinsic Guarantee of Physical Consistency: This invention embeds the physical flux continuity constraint into the cross-domain edge refraction condition of graph message passing, making mass conservation, energy conservation, and flux continuity inherent properties of the message passing rules rather than external penalties. This design fundamentally solves the problem in existing AI-CAE methods where physical constraints rely on soft penalties and may not satisfy physical laws, ensuring that the prediction results naturally satisfy the interface flux continuity condition.

[0037] 6. Reduced hardware resource overhead: This invention eliminates the need for regular mesh homogenization preprocessing and can directly use the original unstructured CAE mesh as input, avoiding the large-scale interpolation calculations and storage overhead required to map unstructured meshes to regular meshes. Compared to traditional CAE solvers, this invention reduces simulation time from hours to seconds while achieving equivalent accuracy.

[0038] 7. Extension of generalization ability: This invention uses one-hot encoding of boundary condition types as fixed inputs to graph nodes, enabling the trained model to receive inputs of arbitrary boundary condition combinations and provide corresponding physical field predictions; material physical property parameters are used as continuous feature inputs to graph nodes, enabling the model to perform zero-sample predictions of new materials not present in the training set.

[0039] 8. Data storage and engineering closed-loop integrity: This invention uses HDF5 format files to uniformly store inference results, SDF geometric features, graph topology, material property parameters and boundary condition parameters, and incorporates all simulation-related data into the same structured file system, providing a complete data foundation for subsequent cloud map rendering, index extraction, parameter iteration and design optimization, and supporting the closed-loop operation of industrial-grade simulation workflow. Attached Figure Description

[0040] Figure 1 This is a diagram showing the overall module architecture of the system of the present invention;

[0041] Figure 2 This is a flowchart of the inference engine mode selection process of this invention. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0044] refer to Figure 1-2This invention provides a general AI-CAE multiphysics simulation system based on a three-layer fusion of SDF, Graph, and Grid, belonging to the field of artificial intelligence-driven industrial simulation technology. Specifically, it involves a rapid multiphysics simulation technology that integrates neural operator methods with geometric topology modeling, applicable to fluid simulation, conjugate heat transfer simulation, high-frequency electromagnetic simulation, and fluid-structure interaction simulation. The system includes: a CAD preprocessing module for performing mesh repair, flow domain extraction, and component partitioning on multi-component CAD models, and binding each component or medium domain to material physical property parameters; an SDF encoding module for generating a segmented multi-medium symbolic distance field across the entire simulation domain, extracting the inner and outer domain masks, distances to walls, distances to medium interfaces, and boundary layer thickness of each node as geometric features; a graph construction module for constructing a graph structure using a set of discrete nodes as graph nodes and adjacency relationships, additionally constructing cross-domain edges for nodes belonging to different medium domains but sharing interfaces, and fusing the unique thermal encoding of geometric features, material physical property parameters, and boundary condition labels into the initial feature vector of each node; and an inference engine. It includes a first inference submodule (aggregating local neighborhood features through graph neural operators), a second inference submodule (projecting graph node features onto an implicit regular grid through a graph encoder, performing global calculations by a frequency domain neural operator, and then back-projecting back to the original graph nodes by a graph decoder), and a third inference submodule (expanding spatial nodes into four-dimensional spatiotemporal nodes on the time axis and constructing spatial and temporal edges, calculating the message passing weights of each edge based on the gradient of the symbolic distance field with respect to spatial and temporal coordinates, and performing iterative propagation until convergence), as well as a mode selection unit that switches between the three modes according to the simulation type and SDF gradient magnitude; and a storage and output module used to write inference results, SDF features, graph topology, and material parameters into a data file.This invention achieves native adaptation to arbitrarily complex CAD geometries and unstructured meshes through SDF implicit geometric encoding and graph topology, avoiding the regular mesh dependency and geometric information loss of pure FNO methods. It accurately distinguishes multi-medium partitions through a segmented multi-medium SDF vector field, using material physical property parameters as continuous node feature inputs, enabling the model to generalize across materials. The mode selection unit adaptively switches between three models based on the SDF gradient magnitude, balancing fine prediction accuracy in local high-gradient regions (boundary layers, medium interfaces) with long-range physical coupling consistency in gradually changing regions. Furthermore, it incorporates the time dimension into a unified graph propagation framework and uses SDF time... - The empty gradient drives the continuous evolution of edge weights, supporting transient simulations of geometric boundary movement without the need for discrete graph reconnection. Combined with temporal edge causal cone constraints, it ensures the physical causality of information propagation. By embedding physical flux continuity constraints into cross-domain edge refraction conditions, physical consistency becomes an intrinsic attribute of message passing rules rather than an external penalty term, fundamentally guaranteeing the continuity of interface flux. The system also avoids large-scale interpolation calculations by directly inputting the original CAE mesh, supports arbitrary combinations of working conditions and zero-sample prediction of unknown materials by using boundary condition one-hot encoding and continuous input of material properties, and uniformly stores all simulation data in structured HDF5 files to support complete industrial closed-loop operation.

[0045] Example 1

[0046] Purpose of implementation:

[0047] This embodiment uses the liquid cooling heat sink of the CPU chip of a data center server as the simulation object to verify the complete implementation process of the present invention in the conjugate heat transfer simulation scenario. Specifically, it verifies: (1) the CAD preprocessing module’s functions of mesh repair, flow domain extraction and component partitioning for multi-component models; (2) the SDF encoding module’s function of generating a segmented SDF vector field and extracting geometric features for multi-medium scenarios containing solid and fluid domains; (3) the graph construction module’s function of constructing cross-domain edges for nodes belonging to different medium domains; (4) the mode selection unit’s function of selecting the first inference submodule based on the SDF gradient magnitude threshold; and (5) the function of the first inference submodule to perform local fine prediction of the boundary layer and interface region through graph neural operators.

[0048] Implementation System:

[0049] This embodiment uses the multiphysics simulation system based on SDF-Graph-Grid fusion provided by the present invention. The specific modules called include: CAD preprocessing module, SDF encoding module, graph construction module, inference engine (including first inference submodule and mode selection unit), and storage and output module.

[0050] Implementation steps:

[0051] Step 1: Import the multi-component CAD model. The simulation object is a silicon-based chip with dimensions of 20mm × 20mm × 2mm, on which a copper microchannel heat sink with dimensions of 20mm × 20mm × 4mm is mounted. The heat sink contains 10 parallel microchannels with a channel width of 0.3mm, a channel height of 2mm, and a fin thickness of 0.2mm. Coolant (deionized water) flows within the microchannels. The CAD model consists of three components: the chip (solid domain), the heat sink substrate and fins (solid domain), and the coolant channels (fluid domain).

[0052] Step two: The CAD preprocessing module performs surface mesh repair, watershed closed domain extraction, and component partitioning. This module removes overlapping surfaces, fills holes, merges coplanar surfaces in the CAD model, identifies coolant channels as closed fluid domains, and marks the chip, heat sink substrate and fins, and coolant channels as different components.

[0053] Step 3: The CAD preprocessing module binds each component to its corresponding material physical property parameters. The chip (silicon) is bound to the following parameters: thermal conductivity 148 W / (m·K), density 2330 kg / m³, specific heat capacity 712 J / (kg·K); the heat sink (copper) is bound to the following parameters: thermal conductivity 401 W / (m·K), density 8930 kg / m³, specific heat capacity 385 J / (kg·K); the coolant (deionized water) is bound to the following parameters: density 997 kg / m³, dynamic viscosity 0.001 Pa·s, thermal conductivity 0.6 W / (m·K), specific heat capacity 4180 J / (kg·K). The system also sets the boundary conditions: the internal heat source of the chip has a heating power of 200W (equivalent to a heat flux density of 50W / cm²), the inlet is a velocity inlet (flow velocity 1m / s, temperature 25°C), the outlet is a pressure outlet (gauge pressure is zero), all solid-liquid interfaces are non-slip wall boundaries and conjugate heat transfer coupling is enabled, the bottom of the chip is an adiabatic boundary, and the outer wall of the heat sink is a convective heat transfer boundary (ambient temperature 25°C, convective heat transfer coefficient 10W / (m²·K)).

[0054] Step four: The SDF encoding module generates a segmented multi-medium SDF vector field. This module calculates the signed distance from each node to the corresponding material interface for each of the three media types, forming the SDF vector field. The three components of the vector field correspond to the silicon, copper, and water media, respectively. For nodes located inside the solid domain, the SDF value is negative; for nodes located inside the fluid domain, the SDF value is positive. This module extracts from the SDF vector field the inner and outer domain mask values ​​(identifying whether the node belongs to the solid or fluid domain), the distance to the nearest wall, the distance to the nearest solid-liquid interface, and the boundary layer thickness value for each node (calculated by inversely using the SDF gradient magnitude; nodes with a gradient magnitude exceeding three times the regional average are marked as boundary layer nodes).

[0055] Step 5: The graph construction module uses the set of discrete nodes as graph nodes and constructs edges based on the adjacency relationships between nodes to form a graph structure. The mesh size is approximately 1.2 million elements, with about 800,000 elements in the fluid domain and about 400,000 elements in the solid domain. This module additionally constructs cross-domain undirected edges for nodes belonging to different medium domains but sharing interfaces, with approximately 20,000 cross-domain edges at the solid-liquid interface. This module fuses the unique thermal encoding of SDF geometric features, material physical property parameters, and boundary condition labels into the initial feature vector of each graph node.

[0056] Step six: The mode selection unit of the inference engine selects the inference submodule based on the simulation type and SDF gradient magnitude. In this embodiment, which is a steady-state simulation, the mode selection unit calculates the SDF gradient magnitude of each node and finds that the SDF gradient magnitude of the nodes near the solid-liquid interface (accounting for about 8% of the total number of nodes) exceeds three times the regional average. Therefore, the first inference submodule is selected.

[0057] Step seven: The first inference submodule takes the graph structure and initial feature vectors as input and aggregates the neighborhood features of each node through the convolutional layers of the graph neural operator. The aggregation method of the graph neural operator layer is as follows: multiply the feature vector of each neighboring node of the given node by the learnable weight matrix, and then multiply by a radial basis function value. The independent variable of the radial basis function is the Euclidean distance between the given node and its neighboring nodes. After summing all products, the updated features of the given node are output through a nonlinear activation function. The network contains a total of 6 graph convolutional layers, each with a hidden dimension of 256. The radial basis function uses a Gaussian kernel with a kernel width of 0.1. After 6 layers of message passing, each node outputs the predicted physical field values, including the temperature scalar field, the velocity vector field, and the pressure scalar field.

[0058] Step 8: The storage and output module writes the inference results, SDF features, graph topology, material parameters, and boundary condition parameters into the data file.

[0059] Implementation results:

[0060] The prediction results of this invention are compared with those of a traditional CAE solver (ANSYS Fluent), using the SIMPLE algorithm and a second-order upwind scheme, and iteratively solved on the same mesh until the residual is less than 10⁻. 6 The calculation results were compared with those of Fluent. In the central region of the chip (the region with the largest temperature gradient), the relative error between the temperature value predicted in this embodiment and the value calculated by Fluent was 1.2%; in the coolant outlet region, the relative error between the predicted temperature and the value calculated by Fluent was 0.8%; and the average relative error of the global temperature field was 0.9%.

[0061] Example 2

[0062] Purpose of implementation:

[0063] This embodiment uses a rectangular microstrip patch antenna in the 5.8 GHz band as the simulation object to verify the implementation process of the present invention in a high-frequency electromagnetic simulation scenario. Specifically, it verifies: (1) the material binding function of the CAD preprocessing module containing the PML absorption layer and the metal-dielectric composite structure; (2) the SDF encoding module generating a four-segment SDF vector field containing the air domain, substrate domain, metal domain and PML absorption domain; (3) the function of the mode selection unit to select the second inference submodule in the steady-state region without large gradients; (4) the complete process of the graph encoder projecting graph node features to the implicit regular grid, the frequency domain neural operator performing global calculation, and the graph decoder back projecting back to the original graph node in the second inference submodule.

[0064] Implementation System:

[0065] This embodiment uses the multiphysics simulation system based on SDF-Graph-Grid fusion provided by the present invention. The specific modules called include: CAD preprocessing module, SDF encoding module, graph construction module, inference engine (including second inference submodule and mode selection unit), and storage and output module.

[0066] Implementation steps:

[0067] Step 1: Import the multi-component CAD model. The simulation object is a rectangular microstrip patch antenna operating at 5.8 GHz. The structure includes: a radiating patch (15.8 mm × 11.6 mm × 0.035 mm, made of copper), a dielectric substrate (30 mm × 30 mm × 1.6 mm, made of Rogers RO4350B, with a relative permittivity εᵣ = 3.48 and a loss tangent tanδ = 0.0037), a ground plane (30 mm × 30 mm × 0.035 mm, made of copper), and an air box surrounding the entire antenna (60 mm × 60 mm × 30 mm, with a PML absorbing layer as the boundary). The feeding method is microstrip line side-feed, with the feed point located at the edge of the radiating patch.

[0068] Step two: The CAD preprocessing module performs surface mesh repair, watershed closure domain extraction, and component partitioning, and binds each component to its corresponding material physical property parameters. Copper binding conductivity is 5.8 × 10⁻⁶. 7 S / m; RO4350B substrate is bonded with a relative permittivity of 3.48 (real part), loss tangent of 0.0037 (imaginary part related), wavenumber k0=2πf / c, and frequency f=5.8GHz; air is bonded with a relative permittivity of 1.0 and a relative permeability of 1.0; the PML absorber layer is bonded with conductivity, which is polynomially gradient from 0 to a maximum value of 0.5S / m, and the permeability is set according to the matching relationship.

[0069] Step 3: The SDF encoding module generates a segmented multi-medium SDF vector field. This module calculates the signed distance from each node to the corresponding material interface for each of the four media types, forming the SDF vector field. The four components of the vector field correspond to the four media: air, substrate, metal, and PML, respectively. From the SDF vector field, this module extracts the inner and outer domain mask values ​​for each node (identifying whether the node belongs to the air domain, substrate domain, metal domain, or PML domain), the distance to the nearest metal surface, the distance to the nearest medium interface, and the boundary layer thickness (used for field attenuation calculation within the PML).

[0070] Step four: The graph construction module uses a set of discrete nodes as graph nodes and constructs edges based on the adjacency relationships between nodes to form a graph structure. The mesh size is approximately 2.5 million elements, including approximately 1.8 million elements for the air cell, approximately 400,000 elements for the substrate, approximately 250,000 elements for the PML layer, and approximately 50,000 elements for the metal patch and ground plane. This module additionally constructs cross-domain undirected edges for nodes belonging to different medium domains but sharing an interface. This module fuses the SDF geometric features, material physical property parameters, and unique thermal encoding of boundary condition labels (port-excited boundaries, PML absorbing boundaries, and ideal electrical conductor boundaries each correspond to different encodings) into the initial feature vector of each graph node.

[0071] Step 5: The mode selection unit of the inference engine selects the inference submodule. In this embodiment, the simulation is a steady-state frequency domain simulation with the excitation frequency fixed at 5.8 GHz. After calculating the SDF gradient magnitude of each node, the mode selection unit finds that there are no nodes with gradient magnitudes exceeding three times the regional average, and the proportion of gradient extremum nodes at the metal-dielectric interface is less than 5%. Therefore, the second inference submodule is selected.

[0072] Step six: The second inference submodule takes the graph structure and initial feature vectors as input. A graph encoder projects the graph node features onto an implicit regular grid. Global computation is then performed on the implicit regular grid using a frequency domain neural operator. Finally, a graph decoder backprojects the computation results back to the original graph nodes. The graph encoder uses a graph attention network encoder, containing three graph attention layers, each with eight attention heads, projecting the graph node features onto an implicit regular grid with a resolution of 64×64×32. The frequency domain neural operator uses a Fourier neural operator. Its convolutional layer performs the following operations: transforms the field data on the implicit regular grid to the frequency domain using a three-dimensional fast Fourier transform; retains the 16 lowest frequency modes in the frequency domain and multiplies them with a learnable frequency domain weight matrix (dimension 16×16×16×256); then transforms it back to the spatial domain using a three-dimensional inverse fast Fourier transform. The graph decoder uses a graph attention network decoder to back-project the calculation results back to the original graph nodes and output the electromagnetic field prediction values ​​of each node, including the electric field vector (containing real and imaginary parts and three directional components) and the magnetic field vector (containing real and imaginary parts and three directional components).

[0073] Step 7: The storage and output module writes the inference results, SDF features, graph topology, material parameters, and boundary condition parameters into the data file.

[0074] Implementation results:

[0075] The prediction results of this invention are compared with the calculation results of a traditional CAE solver (ANSYS HFSS, which uses the finite element method and solves to convergence on the same mesh). At a working frequency of 5.8 GHz, the S predicted in this embodiment... 11 The return loss parameter is -28.5 dB, while the HFSS calculated value is -29.1 dB, with an error of 0.6 dB. The predicted main lobe gain of the radiation pattern is 7.2 dBi, while the HFSS calculated value is 7.4 dBi, with an error of 0.2 dBi. The normalized correlation coefficient between the predicted E-plane radiation pattern and the HFSS calculated value is 0.992. The average prediction error for the real and imaginary parts of the complex field is 3.7%.

[0076] Example 3:

[0077] Purpose of implementation:

[0078] This embodiment takes the vortex-induced vibration of a cylinder as the simulation object, and aims to verify the implementation process of the present invention in the transient fluid-structure interaction (FSI) simulation scenario. Specifically, it verifies: (1) the function of the third inference submodule to expand the three-dimensional spatial nodes into four-dimensional spatiotemporal nodes on the time axis and construct spatial edges and temporal edges; (2) the calculation of the SDF spatiotemporal field value and its four-dimensional gradient with respect to spatial coordinates and time coordinates, and the process of constructing the geodesic tensor by the outer product of the four-dimensional gradient; (3) the method of calculating the message passing weight based on the quadratic value of the edge direction vector and the geodesic tensor; (4) the implementation of the causal cone constraint on the temporal edge, that is, the time weight is reset to zero when the time interval is less than the spatial distance divided by the maximum propagation speed of the physical field; (5) the specific implementation of the line integral on the cross-domain edge using the refraction condition at the medium interface; (6) the complete process of message passing iteration until the maximum value of the relative change of all node messages is less than the preset convergence threshold.

[0079] Implementation System:

[0080] This embodiment uses the multiphysics simulation system based on SDF-Graph-Grid fusion provided by the present invention. The specific modules called include: CAD preprocessing module, SDF encoding module, graph construction module, inference engine (including third inference submodule and mode selection unit), and storage and output module.

[0081] Implementation steps:

[0082] Step 1: Import the multi-component CAD model. The simulation object is an elastically supported cylinder placed in a uniform inflow. The cylinder diameter is D = 0.1 m, the mass per unit length is 1.0 kg / m, the spring stiffness is 100 N / m, and the damping ratio is 0.01. The fluid domain is a rectangular region with dimensions of 20D × 10D, and the cylinder is located 5D from the inlet. The inflow velocity is 1.0 m / s, and the Reynolds number, calculated based on the cylinder diameter and the fluid kinematic viscosity, is approximately 1.0 × 10⁻⁶. 4 .

[0083] Step two: The CAD preprocessing module performs surface mesh repair, watershed closure domain extraction, and component partitioning, binding each component to its corresponding material physical property parameters. The fluid (water) binding density is 1000 kg / m³, dynamic viscosity is 1.0 × 10⁻³ Pa·s; the solid (cylindrical structure) binding density is 7850 kg / m³, elastic modulus is 210 GPa, and Poisson's ratio is 0.3. System boundary conditions are set as follows: the fluid-solid interface is a coupling boundary, the fluid inlet is a velocity inlet (inflow velocity 1.0 m / s, lateral velocity is zero), the outlet is a pressure outlet (gauge pressure is zero), and the upper and lower walls are symmetrical boundaries.

[0084] Step 3: The SDF encoding module generates a segmented multi-medium SDF vector field. This module calculates the signed distance from each node to the corresponding material interface for both media types, forming the SDF vector field. The two components of the vector field correspond to the fluid and solid media, respectively. Since the cylinder will vibrate and displace under the action of fluid forces, the SDF field changes with time, and the SDF value is recalculated at each time step.

[0085] Step four: The graph construction module uses a set of discrete nodes as graph nodes and constructs edges based on the adjacency relationships between nodes to form a graph structure. The mesh size is approximately 500,000 elements (in the two-dimensional case), with approximately 450,000 elements in the fluid domain and approximately 50,000 elements in the solid domain. This module additionally constructs cross-domain undirected edges for nodes belonging to different medium domains but sharing an interface, with approximately 8,000 cross-domain edges at the fluid-solid interface. This module fuses the unique thermal encoding of SDF geometric features, material physical property parameters, and boundary condition labels into the initial feature vector of each graph node.

[0086] Step 5: The mode selection unit of the inference engine selects the inference submodule. In this embodiment, which is a transient simulation, the mode selection unit directly selects the third inference submodule based on the transient type.

[0087] Step six: The third inference submodule expands the two-dimensional spatial nodes on the time axis, generating a three-dimensional spatiotemporal node set (two-dimensional space plus time). The time range is 0 to 10 seconds, with an initial time step of 0.005 seconds, resulting in a total of 2001 time slices. This submodule constructs spatial edges (connecting spatially adjacent nodes within the same time slice) and directed temporal edges (connecting adjacent time slices of the same spatial node, with the direction pointing from the current moment to the next moment) on the three-dimensional spatiotemporal node set.

[0088] Step 7: The third inference submodule calculates the SDF spatiotemporal field value at each 3D spatiotemporal node and its 3D gradient with respect to spatial and temporal coordinates. The 3D gradient contains partial derivatives with respect to two spatial coordinates and one temporal coordinate. The outer product of these 3D gradients forms the geodesic tensor, which is a 3×3 symmetric tensor. For each edge, this submodule calculates the message passing weight based on the quadratic value of the edge direction vector and the geodesic tensor. This weight is the value of a natural exponential function with the negative of the product of the quadratic value and a preset scale parameter, which is set to 1.0. For temporal edges, this submodule determines the relationship between the time interval and the spatial distance: if the time interval is less than the spatial distance divided by the maximum propagation speed of the physical field (the speed of sound in fluids is 1500 m / s), then the weight of the temporal edge is reset to zero.

[0089] Step 8: The third inference submodule performs message passing iterations with the aforementioned weights. Each node's new message is the sum of the neighboring nodes' old messages multiplied by the corresponding edge weight and then multiplied by the geodesic curvature attenuation factor, where the geodesic curvature attenuation factor is the natural exponential function value with the negative integral of the SDF isosurface curvature line along the edge path as the exponent. For cross-domain edges (edges at the fluid-solid interface), the line integral at the medium interface uses the refraction condition: the refractive index of the fluid medium multiplied by the sine of the incident angle equals the refractive index of the solid medium multiplied by the sine of the refraction angle, where the refractive index of the fluid medium is the square root of the fluid density, and the refractive index of the solid medium is the square root of the product of the solid's elastic modulus and density. This iteration is repeated until the maximum value of the relative changes in messages across all nodes is less than the preset convergence threshold of 0.001, converging after 35 iterations. This submodule outputs the predicted physical field values ​​(velocity, pressure of the flow field, and displacement and stress of the solid) for all three-dimensional spatiotemporal nodes.

[0090] Step nine: The storage and output module writes the inference results, SDF features, graph topology, material parameters, and boundary condition parameters into the data file.

[0091] Implementation results:

[0092] The prediction results of this invention are compared with the calculation results of a traditional CAE solver (ANSYS Fluent with TransientStructural bidirectional fluid-structure interaction, time step 0.005s). In this embodiment, the predicted transverse vibration amplitude of the cylinder is 0.032D (dimensionless), while the FSI solver's calculated value is 0.035D, with an error of 8.6%; the predicted vortex shedding frequency is 2.1Hz, while the FSI solver's calculated value is 2.0Hz, with an error of 5.0%; the predicted mean drag coefficient is 1.18, while the FSI solver's calculated value is 1.22, with an error of 3.3%; the root mean square value of the lift coefficient is 0.42, while the FSI solver's calculated value is 0.45, with an error of 6.7%.

[0093] Example 4

[0094] Purpose of implementation:

[0095] This embodiment takes the simulation of turbulent boundary layer of a flat plate as the object, and aims to verify the implementation process of the present invention under the first inference submodule (pure Graph local fine inference mode). Specifically, it verifies: (1) the adaptability of the first inference submodule to the highly anisotropic boundary layer mesh; (2) the anisotropy setting method of the radial basis function kernel in the graph neural operator layer, that is, the normal and tangential directions adopt different kernel widths; (3) the actual operation of the mode selection unit to select the first inference submodule according to the SDF gradient magnitude threshold (the gradient magnitude exceeds three times the average value of the region); (4) the fine prediction accuracy of the local large gradient region (the dimensionless wall distance near the wall is less than 30).

[0096] Implementation System:

[0097] This embodiment uses the multiphysics simulation system based on SDF-Graph-Grid fusion provided by the present invention. The specific modules called include: CAD preprocessing module, SDF encoding module, graph construction module, inference engine (including first inference submodule and mode selection unit), and storage and output module.

[0098] Implementation steps:

[0099] Step 1: Import the multi-component CAD model. The simulation object is a zero-pressure gradient plate placed in a uniform inflow. The plate length L = 1.0m, and the fluid is air (density 1.225kg / m³, dynamic viscosity 1.789×10⁻⁻⁻⁻⁴). 5 The flow velocity is 10 m / s, and the Reynolds number based on the plate length is approximately 5.47 × 10⁻⁶ Pa·s. 5 The simulation domain is the boundary layer region within the range of 0 to 0.05m above the flat surface, with a focus on the fully developed turbulent boundary layer in the range of 0.5m to 0.8m.

[0100] Step two: The CAD preprocessing module performs surface mesh repair, watershed closure domain extraction, and component partitioning, and binds the material physical property parameters of the fluid domain to the air.

[0101] Step 3: The SDF encoding module generates a single-medium SDF field (containing only the fluid domain and solid wall). This module extracts the distance from each node to the wall (wall law coordinates) and the boundary layer thickness from the SDF field (identified by SDF gradient, where the SDF gradient magnitude is extremely large near the wall).

[0102] Step four: The graph construction module uses a set of discrete nodes as graph nodes and constructs edges based on the adjacency relationships between nodes to form the graph structure. The mesh uses prism elements (boundary layer mesh), with refinement at the wall normal height: the first layer mesh height is 5×10⁻ 6 The grid is divided into 60 layers along the normal direction and 200 elements each along the flow and span directions, for a total of approximately 2.4 million elements. The grid exhibits high anisotropy near the wall (regions where the dimensionless wall distance is less than 30), with element aspect ratios exceeding 100:1. This module fuses the one-heat encoding of SDF geometric features, material physical property parameters, and boundary condition labels into the initial feature vector for each graph node.

[0103] Step 5: The mode selection unit of the inference engine selects the inference submodule based on the simulation type and SDF gradient magnitude. In this embodiment, it is a steady-state simulation. After calculating the SDF gradient magnitude of each node, the mode selection unit finds that the SDF gradient magnitude near the wall (the dimensionless wall distance is less than 30, accounting for about 25% of the total number of nodes) is more than three times the regional average, and the SDF gradient magnitude at the wall can reach 50 times the regional average. Therefore, the first inference submodule is selected.

[0104] Step six: The first inference submodule takes the graph structure and initial feature vectors as input and aggregates the neighborhood features of each node through convolutional layers of graph neural operators. This submodule uses anisotropic radial basis function kernel settings: the normal kernel width is 0.1 times the height of the first layer grid, and the tangential kernel width is the flow grid spacing. The network contains a total of 8 graph convolutional layers, each with a hidden dimension of 512. After 8 layers of message passing, each node outputs predicted physical field values, including flow velocity, normal velocity, turbulent kinetic energy, and dissipation rate.

[0105] Step 7: The storage and output module writes the inference results, SDF features, graph topology, material parameters, and boundary condition parameters into the data file.

[0106] Implementation results:

[0107] The prediction results of this invention were compared with direct numerical simulation data (taken from the standard zero-pressure gradient turbulent boundary layer database). At the x=0.6m section, the average flow velocity profile predicted in this embodiment showed extremely high agreement with the direct numerical simulation data: in the logarithmic law region (dimensionless wall distance greater than 30), the velocity prediction error was less than 1%; in the viscous sublayer (dimensionless wall distance less than 5), the velocity prediction error was less than 2%; and over the entire boundary layer thickness, the average relative error was 1.8%. The prediction error of turbulent kinetic energy was approximately 5% in the near-wall region (dimensionless wall distance less than 50) and less than 3% in the outer region.

[0108] Example 5

[0109] Purpose of implementation:

[0110] This embodiment takes the shock tube problem (Sod shock tube) as the simulation object, and aims to verify the core mechanism of the third inference submodule of the present invention in the transient scenario with strong discontinuity (shock wave). Specifically, it verifies: (1) the complete mathematical process of the time partial derivative of the SDF spacetime field generating a non-zero value as the geometric boundary (shock wave) moves, and being transferred to the edge weight through the geodesic tensor; (2) the operation mode of the edge weight continuously evolving as the geometric boundary moves without the need for discrete reconnection of the graph topology; (3) the physical effect of the edge weight approaching zero due to the maximum SDF spatial gradient at the moving shock surface, and the message being unable to cross the strong discontinuity surface; (4) the implementation of causal cone constraints on the temporal edge.

[0111] Implementation System:

[0112] This embodiment uses the multiphysics simulation system based on SDF-Graph-Grid fusion provided by the present invention. The specific modules called include: CAD preprocessing module, SDF encoding module, graph construction module, inference engine (including third inference submodule and mode selection unit), and storage and output module.

[0113] Implementation steps:

[0114] Step 1: Establish the simulation object and physical settings. The Sod shock tube is a one-dimensional Riemann problem with a computational domain of 0 to 1 m. Initially, a diaphragm is placed at x = 0.5 m. To the left of the diaphragm is high-pressure gas (density 1.0 kg / m³, pressure 1.0 Pa, velocity zero), and to the right is low-pressure gas (density 0.125 kg / m³, pressure 0.1 Pa, velocity zero). The diaphragm ruptures at time zero, forming a shock wave propagating to the right, an expansion wave propagating to the left, and a contact discontinuity. The simulation time is 0 to 0.2 s.

[0115] Step 2: The SDF encoding module generates the SDF time-space field. The SDF time-space field is defined as the signed distance from the spatial point coordinates to the shock wave position, which is the spatial point coordinates minus the shock wave position coordinates. The shock wave position moves to the right with time, and the shock wave position is approximately equal to 0.5m plus the shock wave velocity multiplied by time. The shock wave velocity is approximately 1.0m / s. Therefore, the partial derivative of the SDF with respect to time is the negative shock wave velocity, i.e., a non-zero value.

[0116] Step 3: The graph construction module uses a set of discrete nodes as graph nodes and constructs edges based on the adjacency relationships between nodes to form a graph structure. The number of spatial nodes is 1000, evenly distributed in the interval from 0 to 1m.

[0117] Step four: The mode selection unit of the inference engine selects the inference submodule. In this embodiment, which is a transient simulation, the mode selection unit directly selects the third inference submodule.

[0118] Step 5: The third inference submodule expands the one-dimensional spatial nodes on the time axis, generating a two-dimensional spatiotemporal node set (one-dimensional space plus time). There are 200 time slices, a total time of 0.2 seconds, and a time step of 0.001 seconds. This submodule constructs spatial edges (connecting adjacent spatial nodes within the same time slice) and directed temporal edges (connecting adjacent time slices of the same spatial node) on the two-dimensional spatiotemporal node set.

[0119] Step Six: The third inference submodule calculates the SDF spatiotemporal field value and its two-dimensional gradient with respect to spatial and temporal coordinates at each two-dimensional spatiotemporal node. The two-dimensional gradient contains the partial derivatives with respect to spatial and temporal coordinates, and the outer product of these two-dimensional gradients constitutes the geodesic tensor. For each edge, this submodule calculates the message propagation weight based on the quadratic form of the edge direction vector and the geodesic tensor. Key observation: As the shock wave moves to the right, at t=0.05s the shock wave is located at approximately 0.55m, at t=0.10s the shock wave is located at approximately 0.60m, and at t=0.15s the shock wave is located at approximately 0.65m. The edge weights evolve continuously—at the shock wave location, the SDF spatial gradient is extremely large, the edge weights approach zero, and the message cannot cross the shock wave surface; behind the shock wave (in the expansion wave region), the SDF spatial gradient is small, the edge weights are large, and the message propagates normally; in front of the shock wave (in the undisturbed region), the SDF spatial gradient is zero, the edge weights are 1.0, and the message propagates without attenuation. Throughout the process, the graph topology remains unchanged, with only the edge weights changing continuously.

[0120] Step seven: The third inference submodule performs message passing iterations, with a convergence threshold of 0.001, and converges after 25 iterations. This submodule outputs the predicted physical field values ​​(density, velocity, pressure) for all two-dimensional spatiotemporal nodes.

[0121] Step 8: The storage and output module writes the inference results, SDF features, graph topology, material parameters, and boundary condition parameters into the data file.

[0122] Implementation results:

[0123] The prediction results of this invention are compared with the exact solution of the Riemann problem. At t=0.1s: the predicted shock wave location in this embodiment is 0.601m, while the exact solution is 0.600m, with an error of 0.17%; the predicted density ratio before and after the shock wave is 4.02, while the exact solution is 4.0, with an error of 0.5%; the relative error between the pressure distribution in the expansion wave region and the exact solution is less than 2%. The correlation coefficient between the spatiotemporal contour map of the entire transient process (0 to 0.2s) and the exact solution is 0.994.

[0124] Comparative Example

[0125] Purpose of implementation:

[0126] This comparative example uses the pure Fourier neural operator (FNO) method to simulate the conjugate heat transfer simulation problem of the chip liquid cooling heat sink in Example 1. The aim is to highlight the significant progress of the present invention in the following aspects by comparing it with the technical solution of the present invention: (1) the geometric expression accuracy advantage of SDF implicit geometric coding over regular mesh mask approximation; (2) the native adaptability of Graph topology to unstructured CAE mesh over regular mesh; (3) the accurate differentiation capability of segmented multi-medium SDF vector field over single mask matrix for multi-medium partitions; (4) the advantage of cross-domain edge refraction conditions in terms of physical consistency over data-driven methods without physical constraints.

[0127] Implementation System:

[0128] This comparative example uses a pure Fourier neural operator method to construct the simulation system, and does not include the geometric repair function, SDF encoding module, graph construction module, three-model adaptive inference engine, and cross-domain edge physical constraint mechanism of the CAD preprocessing module of this invention. The system only includes the following components: a data preprocessing unit (for mapping unstructured mesh data to regular mesh), a Fourier neural operator model (containing 4 Fourier layers), and a result postprocessing unit (for mapping regular mesh results back to unstructured mesh).

[0129] Implementation steps:

[0130] Step 1: Select a benchmark. This comparative example uses the conjugate heat transfer simulation of the chip liquid-cooled heat sink in Example 1 as the benchmark, employing the exact same simulation object, CAD model, material physical property parameters, and boundary condition settings as Example 1. The simulation object is a silicon-based chip with dimensions of 20mm × 20mm × 2mm, topped with a copper microchannel heat sink, and the coolant is deionized water. Material parameters are: Silicon – thermal conductivity 148 W / (m·K), density 2330 kg / m³, specific heat capacity 712 J / (kg·K); Copper – thermal conductivity 401 W / (m·K), density 8930 kg / m³, specific heat capacity 385 J / (kg·K); Deionized water – density 997 kg / m³, dynamic viscosity 0.001 Pa·s, thermal conductivity 0.6 W / (m·K), specific heat capacity 4180 J / (kg·K). The boundary conditions are: 200W internal heat source of the chip, inlet flow rate of 1m / s, temperature of 25°C, and zero outlet gauge pressure.

[0131] Step two, perform mesh homogenization preprocessing. Since the pure Fourier neural operator method only supports regular structured mesh input, the unstructured CAE mesh of the chip heat sink needs to be preprocessed as follows: First, forcibly map the entire simulation domain (including the solid domain and the fluid domain) to a uniform regular rectangular mesh with a resolution of 256×256×128; Second, mark the boundaries of the solid domain and the fluid domain with a mask matrix on the regular mesh using the nearest neighbor interpolation method, where 0 represents fluid and 1 represents solid; Third, define the material physical properties (thermal conductivity, viscosity, etc.) as functions of spatial location and assign values ​​at each grid point of the regular mesh.

[0132] Step 3: Construct the Fourier Neural Operator (FNO) model. An FNO model with four Fourier layers is established, each layer retaining 12 lowest-frequency modes, with a hidden dimension of 64. The input to this model is the material property field and boundary condition encoding on a regular mesh, and the output is the temperature and velocity fields on the regular mesh.

[0133] Step 4, Model Training. Using the same training dataset as in Example 1 (conjugate heat transfer simulation results generated by ANSYS Fluent on the same grid, containing 10,000 sets of temperature and velocity fields under different operating parameters), the FNO model is trained until the loss function converges, with mean squared error as the loss function.

[0134] Step 5: Perform inference prediction. Input the test conditions from Example 1 (chip heating power 200W, inlet flow velocity 1m / s) into the trained FNO model, perform forward inference, and output the predicted values ​​of temperature field and velocity field on the regular grid.

[0135] Step Six, Post-processing of Results. The prediction results on the regular grid are mapped back to the original unstructured CAE grid using the nearest neighbor interpolation method, so as to compare them with the prediction results of the present invention and the Fluent calculation values ​​in Example 1. Implementation Results:

[0136] Under the same simulation conditions (chip heating power 200W, inlet flow velocity 1m / s), the comparison data between the prediction results of the pure Fourier neural operator method and the Fluent calculation values ​​are as follows:

[0137] Chip center temperature prediction error: 18.7% in this comparative example, while 1.2% in Example 1 of the present invention.

[0138] The error in predicting the solid-liquid interface temperature is 25.3% in this comparative example and 1.5% in Example 1 of the present invention.

[0139] Coolant outlet temperature prediction error: 12.1% in this comparative example, while 0.8% in Example 1 of the present invention.

[0140] Prediction error of velocity field near the wall: 34.6% in this comparative example, while 2.1% in Example 1 of the present invention.

[0141] Does mesh homogenization preprocessing need to be performed? Yes in this comparative example (takes about 15 minutes), no in this invention.

[0142] Multi-media partitioning is supported: This comparison shows no (mask approximation required), while this invention shows yes (native support).

[0143] The mechanism by which the pure Fourier neural operator method fails in complex geometric multiphysics simulations is analyzed as follows:

[0144] First, the geometric distortion mechanism—when unstructured CAE meshes are forcibly mapped to regular rectangular meshes, the complex boundaries of the solid and fluid domains (sharp corners of microchannel fins, curved surfaces of solid-liquid interfaces) cannot be accurately represented on the regular mesh. There is a misalignment between the grid points of the regular mesh and the original geometric boundaries. This misalignment introduces significant spatial discretization errors near the boundaries, resulting in a severe loss of geometric information.

[0145] Second, the ambiguity mechanism of the material interface—the regular grid of the Fourier neural operator can only approximate the material boundary at the solid-liquid interface through 0 and 1 transitions in the mask matrix, while the actual material boundary is a sharp geometric surface. The mask approximation introduces a numerical diffusion effect near the interface, causing severe distortion in the predicted temperature and velocity fields at the interface. In Example 1, the temperature prediction error at the solid-liquid interface of this invention is 1.5%, while in the comparative example it is as high as 25.3%. This difference directly reflects the combined effect of geometric distortion and interface ambiguity.

[0146] Third, the anisotropic mesh mismatch mechanism—near the boundary layer, a finer mesh needs to be refined in the wall normal direction. The first layer mesh is only 5 μm high, with a spanwise mesh size of about 0.5 mm and an aspect ratio exceeding 100:1. Fourier neural operators require uniform meshes in all directions; if a uniform mesh is used to achieve the same normal resolution, the total number of meshes will exceed 10. 9 The magnitude is far beyond actual computing power. Limited by available computing resources, the 256×256×128 regular mesh used in this comparative example can only provide approximately 10⁻⁻⁶ in the wall normal direction. 4 A resolution on the order of meters is far from sufficient to resolve the velocity gradient within the boundary layer, which is the fundamental reason why the velocity field prediction error near the wall is as high as 34.6%.

[0147] Fourth, the lack of physical constraints—the pure Fourier neural operator method relies solely on data-driven approaches and does not embed any physical consistency constraints (such as the conservation of heat flux at the solid-liquid interface). During training, this method only minimizes the data fitting loss and cannot guarantee that the prediction results satisfy fundamental physical laws. Near the solid-liquid interface, due to the numerical diffusion of the mask approximation and the lack of flux conservation constraints, heat flux discontinuities are severe, and the temperature field prediction violates the basic requirements of Fourier's law of heat conduction.

[0148] Comparing Examples 1-5 with Comparative Example 1, the comparison results of the above five examples and one comparative example fully demonstrate the comprehensive superiority of the present invention over the pure FNO method. In Example 1 (conjugate heat transfer), the present invention significantly outperforms the pure FNO method: the chip center temperature prediction error is 1.2% compared to 18.7% for pure FNO; the solid-liquid interface temperature prediction error is 1.5% compared to 25.3% for pure FNO; the coolant outlet temperature prediction error is 0.8% compared to 12.1% for pure FNO; and the velocity field prediction error near the wall is 2.1% compared to a high 34.6% for pure FNO. In Example 2 (high-frequency electromagnetic simulation), the present invention simulates the S-wave velocity field of a 5.8GHz microstrip patch antenna. 11The parameter prediction error is only 0.6 dB, the pattern gain error is only 0.2 dBi, and the E-plane pattern correlation coefficient reaches 0.992. In contrast, the pure FNO method, due to its inability to accurately represent the gradual conductivity distribution of the PML absorber layer and the surface current singularity of the metal boundary, typically exhibits prediction errors exceeding 15% in electromagnetic simulations. In Example 3 (Fluid-Structure Interaction (FSI), the predicted transverse vibration amplitude error of the cylinder is 8.6%, the eddy current frequency error is 5.0%, the drag coefficient error is 3.3%, and the lift coefficient error is 6.7%. The pure FNO method, unable to handle the SDF spatiotemporal field evolution caused by moving boundaries, cannot effectively implement fluid-structure interaction simulations. In Example 4 (Turbulent Boundary Layer), the velocity profile prediction error of the present invention on highly anisotropic boundary layer meshes (aspect ratio exceeding 100:1) is less than 1% in the logarithmic law region and less than 2% in the viscous sublayer. The pure FNO method, requiring isotropic meshes, cannot achieve sufficient normal resolution while maintaining computational feasibility, and its near-wall prediction error is typically above 20%. In Example 5 (moving shock wave), the shock wave position error predicted by the present invention is 0.17%, the density ratio error is 0.5%, and the spatiotemporal cloud map correlation coefficient is 0.994. The pure FNO method typically has an error of more than 10% in the problem of moving strong discontinuities because it cannot form a sufficiently sharp discontinuity resolution at the shock wave surface.

[0149] The failure mechanism of the pure FNO method revealed by the comparative examples contrasts sharply with the successful implementation of the embodiments of the present invention. The fundamental difference between the two lies in the following: The present invention uses SDF implicit geometric coding to accurately represent complex CAD boundaries (the microchannel heat sink in Embodiment 1, the antenna structure in Embodiment 2, the cylindrical surface in Embodiment 3, the flat plate wall in Embodiment 4, and the moving shock surface in Embodiment 5 are all accurately represented), preserving all the accuracy of the original geometry without the need for regular mesh homogenization preprocessing; it uses Graph topology to natively adapt unstructured meshes (1.2 million unstructured mesh in Embodiment 1, 2.5 million mesh in Embodiment 2, 0.5 million mesh in Embodiment 3, 2.4 million height anisotropic mesh in Embodiment 4, and 100 million height anisotropic mesh in Embodiment 5). All 0-node one-dimensional meshes are directly used as graph structure inputs, preserving all topological and geometric information of the original mesh without interpolation mapping. A piecewise multi-medium SDF vector field is used to accurately distinguish multi-medium partitions (the three-medium system in Example 1, the four-medium system in Example 2, and the two-medium system in Example 3 are all accurately distinguished), precisely identifying the boundary positions and interface topologies of each medium domain without mask approximation. Cross-domain edge refraction conditions intrinsically guarantee physical flux continuity, treating physical consistency constraints as inherent properties of graph propagation rules rather than external penalties (continuous heat flux at the solid-liquid interface in Example 1, tangential continuity of the medium interface field in Example 2, force balance at the fluid-solid interface in Example 3, and wall boundary conditions in Example 4 are all intrinsically guaranteed). These four core differences collectively determine that this invention can achieve accuracy comparable to traditional CAE in industrial-grade complex geometric multiphysics simulations, while the prediction error of the pure FNO method in the same scenario exceeds that of this invention by more than an order of magnitude.

[0150] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0151] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A general AI-CAE multiphysics simulation system based on SDF-Graph-Grid three-layer fusion, characterized in that, include: The CAD preprocessing module is used to perform mesh repair, watershed extraction and component partitioning on multi-component CAD models, and bind each component or each medium domain to the corresponding material physical property parameters; The SDF encoding module is used to generate a segmented multi-media symbolic distance field and extract the geometric features of each node from the symbolic distance field. The graph construction module is used to construct edges based on the adjacency relationships between discrete nodes, forming a graph structure. It also constructs cross-domain edges for nodes belonging to different medium domains but sharing an interface. The module also integrates the geometric features, material physical property parameters, and boundary condition labels into the initial feature vector of each node. The inference engine includes a first inference submodule, a second inference submodule, a third inference submodule, and a mode selection unit; The first inference submodule is used to take the graph structure and the initial feature vector as input, aggregate the neighborhood features of the nodes through graph neural operators, and output the physical field prediction value of each node. The second inference submodule is used to take the graph structure and the initial feature vector as input, project the graph node features onto the implicit regular grid through the graph encoder, perform global calculation on the implicit regular grid through the frequency domain neural operator, and then project the calculation result back to the original graph node through the graph decoder to output the physical field prediction value of each node. The third inference submodule is used to expand the three-dimensional spatial nodes into four-dimensional spatiotemporal nodes on the time axis, construct a four-dimensional spatiotemporal graph containing spatial edges and temporal edges, calculate the message passing weight of each edge according to the gradient of the symbolic distance field and perform iterative propagation, and output the physical field prediction value of the four-dimensional spatiotemporal node. The mode selection unit is used to select the first inference submodule, the second inference submodule, or the third inference submodule to perform inference based on the simulation type and geometric features. The storage and output module is used to write inference results, symbolic distance field features, graph topology, and material parameters into a data file.

2. The system according to claim 1, characterized in that, The selection rules for the mode selection unit are as follows: when the simulation type is steady state and there is a node in the graph structure whose sign distance field gradient magnitude is greater than three times the average gradient magnitude of all nodes in the region where the node is located, the first inference submodule is selected; when the simulation type is steady state and there is no such node, the second inference submodule is selected; when the simulation type is transient, the third inference submodule is selected.

3. The system according to claim 1, characterized in that, The geometric features extracted by the SDF encoding module include inner and outer domain masks, distances to walls, distances to medium interfaces, and boundary layer thickness.

4. The system according to claim 1, characterized in that, The graph neural operator is a graph neural operator network. The way it aggregates the neighborhood features of each node is as follows: multiply the feature vector of each neighboring node of the node by the learnable weight matrix, and then multiply by a radial basis function value. The independent variable of the radial basis function is the Euclidean distance between the node and its neighboring nodes. After summing all the products, the updated features of the node are output through a nonlinear activation function.

5. The system according to claim 1, characterized in that, The graph encoder is a graph attention network encoder or a graph convolutional network encoder, used to project graph node features onto an implicit regular grid; the graph decoder is a graph attention network decoder or a graph convolutional network decoder, used to back-project the computation results on the implicit regular grid back to the original graph nodes.

6. The system according to claim 1, characterized in that, The frequency domain neural operator is a Fourier neural operator, and its convolutional layer performs the following operations: transforming the field data on the implicit regular grid to the frequency domain via Fourier transform, multiplying it with a learnable frequency domain weight matrix in the frequency domain, and then transforming it back to the spatial domain via inverse Fourier transform.

7. The system according to claim 1, characterized in that, The specific method by which the third inference submodule performs four-dimensional spacetime graph propagation inference is as follows: Expand the three-dimensional spatial nodes on the time axis to generate a four-dimensional spatiotemporal node set, and construct spatial edges and temporal edges on this set; Calculate the symbolic distance field value at each four-dimensional node and its four-dimensional gradient with respect to spatial and temporal coordinates. The geodesic tensor is formed by the outer product of these four-dimensional gradients. For each edge, the message passing weight is calculated based on the quadratic value of the edge direction vector and the geodesic tensor. The weight is the value of a natural exponential function with the negative value of the product of the quadratic value and the preset scale parameter as the exponent. For temporal edges, if the time interval of the edge is less than the spatial distance divided by the maximum propagation speed of the physical field, the weight is set to zero. The message passing iteration is performed with the weights, and the new message of each node is the sum of the old messages of the neighboring nodes multiplied by the corresponding edge weight and the geodesic curvature attenuation factor, wherein the geodesic curvature attenuation factor is the value of the natural exponential function with the negative value of the integral of the sign distance field isosurface curvature line along the edge path as the exponent. For cross-domain edges, the line integral adopts the refraction condition at the medium interface, that is, the ratio of the sine values ​​of the refractive indices of the two media is equal to the ratio of the sine values ​​of the incident angle and the sine value of the refraction angle, and the refractive index is obtained by mapping the material physical property parameters of the medium. Repeat the iteration until the maximum value of the relative change of messages of all nodes is less than the preset convergence threshold, and output the physical field prediction values ​​of all four-dimensional spatiotemporal nodes.

8. The system according to claim 7, characterized in that, The preset scale parameter ranges from 0.01 to 10.0, and the preset convergence threshold ranges from 0.0001 to 0.

01. In the third inference submodule, the time step is adjusted as follows: in each time step, the absolute value of the time partial derivative of the symbol distance field of each node is calculated, and the maximum value is taken. If the maximum value is greater than the first threshold, the current time step is reduced; if the maximum value is less than the second threshold, the current time step is expanded. The first threshold is 0.1, and the second threshold is 0.

01.

9. The system according to claim 1, characterized in that, The boundary condition labels include at least one of the following types: velocity inlet boundary, pressure outlet boundary, no-slip wall boundary, symmetric boundary, periodic boundary, heat flux boundary, convective heat transfer boundary, radiation boundary, lumped port excitation boundary, and wave port excitation boundary; the material physical property parameters include: fluid density, dynamic viscosity, and thermal conductivity; solid thermal conductivity, elastic modulus, and Poisson's ratio; electromagnetic medium relative permittivity, loss tangent, and wavenumber; and PML absorber layer electrical conductivity and magnetic permeability.

10. The system according to claim 1, characterized in that, The multiphysics simulation is one of the following types: fluid simulation, whose governing equations are the Navier-Stokes equations, and the viscosity and thermal conductivity in the equations are temperature-dependent parameters; The conjugate heat transfer simulation uses the Navier-Stokes equations coupled with the energy equations as its governing equations, and the thermal conductivity of the solid is a temperature-dependent parameter. The high-frequency electromagnetic simulation uses Maxwell's equations as its governing equations, and the dielectric constant and permeability are frequency-dependent parameters. The fluid-structure interaction simulation uses the Navier-Stokes equations coupled with the solid mechanics equations as its governing equations, and the elastic modulus of the solid is a temperature-dependent parameter.