An ai finite element solver intelligent acceleration method and system based on physical prior
By optimizing the finite element solver using a physics-based AI model and utilizing the residuals of physical equations as hard constraints, a training dataset is constructed and the solver parameters are adaptively adjusted. This solves the problems of low iterative convergence efficiency and high manual intervention in traditional finite element solvers, achieving efficient and stable simulation results and adaptive optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA THREE GORGES CORPORATION
- Filing Date
- 2026-05-15
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional finite element solvers have low iterative convergence efficiency, rely on human experience, and have high costs for manual intervention. Existing acceleration methods are greatly affected by the amount of data, making them difficult to adapt to complex working conditions. Furthermore, they do not incorporate prior physical knowledge, resulting in insufficient simulation efficiency and accuracy.
The AI model is built based on physical priors, and the residuals of physical equations are used as hard constraints. A loss function is embedded, a training dataset is constructed, and the model parameters are optimized. The solver parameters are adaptively adjusted to form a closed-loop optimization mechanism to ensure that the simulation results meet the physical laws.
It significantly improves the efficiency of finite element simulation, adapts to complex engineering scenarios, reduces manual intervention, ensures that simulation results conform to physical laws, adapts to different materials and working conditions, and its solution efficiency and stability continue to improve in long-term use.
Smart Images
Figure CN122242619A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of finite element simulation technology, and in particular relates to an intelligent acceleration method and system for AI finite element solvers based on physical priors. Background Technology
[0002] Finite element method (FEM) simulation technology is a core technology for structural analysis and operational simulation in fields such as geotechnical engineering and underground engineering. The efficiency of its solution process directly determines the simulation cycle and engineering design efficiency. Whether it's single-physics simulation or multi-physics coupled simulation, the solution for each independent physics field relies on the iterative calculation of the finite element solver. The iterative process of traditional finite element solvers suffers from the following core technical challenges: (1) Low iterative convergence efficiency: Under different working conditions, the iterative algorithm, preconditioner selection, and convergence criterion setting of a single physical field solver all rely on the engineer's manual experience, which can easily lead to iterative divergence and slow convergence speed, resulting in a significant extension of the single field solution cycle and thus affecting the overall simulation efficiency. (2) High cost of manual intervention: Engineers need to have rich experience in finite element solution in order to select appropriate preconditioners and iterative algorithms according to specific single-field working conditions. For novice engineers, it is difficult to quickly adapt to complex working conditions, and the manual debugging process is time-consuming and labor-intensive. (3) Existing acceleration methods have limitations: Most of the finite element acceleration methods on the market rely on data fitting proxy models, which require a large amount of finite element simulation to obtain training data. Not only is the preliminary preparation work cumbersome, but the accuracy of the proxy model is greatly affected by the amount of data, resulting in poor versatility and difficulty in adapting to the varied heterogeneous single-field working conditions in geotechnical engineering. It is impossible to balance acceleration efficiency and simulation accuracy. At the same time, existing acceleration methods for solvers do not incorporate physical prior knowledge, which can easily lead to non-physical iterative results.
[0003] Existing acceleration methods for multi-physics coupling only focus on the interaction and iteration between multiple fields, without optimizing the iteration efficiency of the individual physics solver itself. This invention focuses on "kernel optimization of the individual physics solver", which can accelerate both single-field simulation and single-field solution in multi-field coupling, and is completely different from the acceleration of multi-field coupling decoupling. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an intelligent acceleration method and system for AI finite element solvers based on physical priors.
[0005] Firstly, the method includes, Determine the physical equations corresponding to a single physical field in the target finite element simulation, and extract the physical equation residuals as hard constraints for the AI model based on the physical equations corresponding to a single physical field. Construct an AI model and embed the physical equation residuals as constraints into the loss function of the AI model to obtain a physical constraint loss function; A training dataset is constructed based on this single-field physical prior knowledge; The AI model is trained using the training dataset, and the model parameters are optimized to ensure that the single-field solver parameters output by the AI model can satisfy the residual constraints of the single-field physical equation. The trained AI model is embedded into the existing finite element solver kernel, and the optimal solver parameters are obtained for single-field simulation of the target. After a single-field solution is completed, the AI model adaptively adjusts the output solver parameters based on the physical equation residuals of the solution results, optimizing the efficiency of subsequent simulations of that single field and forming a closed-loop optimization.
[0006] Furthermore, the process of determining the physical equations corresponding to individual physical fields in the target finite element simulation, and extracting the residuals of these physical equations as hard constraints for the AI model based on these equations, specifically includes: The target physical field type is clearly defined; wherein, the target physical field type includes stress-displacement field, seepage field, temperature field, electric field, magnetic field and flow field; Based on the governing equations, constitutive relations, geometric equations and boundary conditions of the target physical field, the physical governing equations of the continuum are established and discretized into a finite element numerical scheme to obtain the equilibrium equations of a single-field finite element system. Based on the equilibrium equations of the finite element system, the physical equation residuals are defined.
[0007] Furthermore, the physical equation residuals include equilibrium equation residuals, constitutive equation residuals, geometric equation residuals, boundary condition residuals, and conservation law residuals.
[0008] Furthermore, the training dataset includes input features, output labels, and physical constraints; The input features are single-field physical parameters, matrix features, and working condition features under different geotechnical conditions, materials, grids, and loads. The output labels are the optimal solver parameter combinations and convergence performance indices for the corresponding working conditions. The physical constraints are the physical equation residuals, the degree of satisfaction of conservation laws, and physical rationality indicators.
[0009] Furthermore, the optimization of the efficiency of subsequent simulations in this single field, forming a closed-loop optimization, specifically includes: Extract the residuals, iteration steps, convergence time, and accuracy metrics of the currently solved physical equations and feed them back to the AI model; The AI model adaptively updates weights based on real-time feedback information, dynamically optimizes parameter output strategies for subsequent operating conditions, and forms a closed-loop optimization mechanism of "perception-decision-execution-feedback".
[0010] Secondly, the system includes: a hard constraint extraction module, an AI model construction module, a training set construction module, a model parameter optimization module, an optimal parameter acquisition module, and an adaptive adjustment module; The hard constraint extraction module is used to determine the physical equation corresponding to a single physical field in the target finite element simulation, and extract the physical equation residual as the hard constraint condition of the AI model based on the physical equation corresponding to the single physical field. The AI model building module is used to build an AI model, embedding the physical equation residuals as constraints into the loss function of the AI model to obtain a physical constraint loss function. The training set construction module is used to construct a training dataset based on the prior physical knowledge of this single field. The model parameter optimization module is used to train the AI model using the training dataset, optimize the model parameters, and ensure that the single-field solver parameters output by the AI model can meet the residual constraints of the single-field physical equation. The optimal parameter acquisition module is used to embed the trained AI model into the existing finite element solver kernel and obtain the optimal solver parameters for the target single-field simulation. The adaptive adjustment module is used to adaptively adjust the output solver parameters of the AI model based on the physical equation residuals of the solution results after a single field is solved, thereby optimizing the efficiency of subsequent simulations of that single field and forming a closed-loop optimization.
[0011] Furthermore, the hard constraint extraction module is specifically used for: The target physical field type is clearly defined; wherein, the target physical field type includes stress-displacement field, seepage field, temperature field, electric field, magnetic field and flow field; Based on the governing equations, constitutive relations, geometric equations and boundary conditions of the target physical field, the physical governing equations of the continuum are established and discretized into a finite element numerical scheme to obtain the equilibrium equations of a single-field finite element system. Based on the equilibrium equations of the finite element system, the physical equation residuals are defined.
[0012] Furthermore, the physical equation residuals include equilibrium equation residuals, constitutive equation residuals, geometric equation residuals, boundary condition residuals, and conservation law residuals.
[0013] Furthermore, the training dataset includes input features, output labels, and physical constraints; The input features are single-field physical parameters, matrix features, and working condition features under different geotechnical conditions, materials, grids, and loads. The output labels are the optimal solver parameter combinations and convergence performance indices for the corresponding working conditions. The physical constraints are the physical equation residuals, the degree of satisfaction of conservation laws, and physical rationality indicators.
[0014] Furthermore, the adaptive adjustment module is specifically used for, Extract the residuals, iteration steps, convergence time, and accuracy metrics of the currently solved physical equations and feed them back to the AI model; The AI model adaptively updates weights based on real-time feedback information, dynamically optimizes parameter output strategies for subsequent operating conditions, and forms a closed-loop optimization mechanism of "perception-decision-execution-feedback".
[0015] Thirdly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above-described intelligent acceleration methods for AI finite element solvers based on physical priors.
[0016] Fourthly, an electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus. Memory, used to store computer programs; When a processor executes a program stored in memory, it implements any of the steps described above for the intelligent acceleration method of the AI finite element solver based on physical priors.
[0017] Compared with the prior art, the present invention has the following advantages: 1. This invention proposes to embed the AI model into the solver kernel, quickly output the optimal combination of solver parameters, eliminate the repeated trial and error process, significantly shorten the calculation time of finite element simulation, significantly improve the solution efficiency, and adapt to complex engineering scenarios with large scale and high mesh density.
[0018] 2. This invention uses the residuals of physical equations as hard constraints to embed loss functions, covering core physical rules such as equilibrium equations, constitutive relations, and boundary conditions. It follows physical laws throughout the entire process from model training to inference, eliminating the problem of AI output parameters violating physical laws from the root, and ensuring that single-field simulation results such as stress-displacement, seepage, and temperature satisfy conservation laws and boundary constraints.
[0019] 3. Relying on real-time solution feedback, the system automatically extracts metrics such as residuals, iteration steps, and convergence time, dynamically updates AI model weights, and adaptively adjusts solver parameter output strategies, constructing a complete closed loop of "perception-decision-execution-feedback." As the number of simulation conditions increases, the model continuously evolves and optimizes, adapting to different materials, meshes, loads, and geotechnical conditions. Over long-term use, its solution efficiency and stability continuously improve, requiring no secondary manual intervention, thus reducing the barrier to entry and maintenance costs.
[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 The flowchart of an AI finite element solver intelligent acceleration method based on physical priors is shown.
[0023] Figure 2 A schematic diagram of an AI finite element solver intelligent acceleration module based on physical priors is shown in the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] like Figure 1 As shown, this invention proposes an intelligent acceleration method for AI finite element solvers based on physical priors. By embedding physical mechanisms into the AI model, it achieves intelligent optimization of the core parameters of a single physics field solver without modifying the simulation model and mesh. This method balances acceleration efficiency and simulation accuracy, reduces manual intervention costs, adapts to complex single-field conditions such as geotechnical engineering, and is compatible with single-field solution stages in multi-field coupling. The steps include... S1. Determine the physical equation corresponding to a single physical field in the target finite element simulation, and extract the residual of the physical equation as the hard constraint condition of the AI model.
[0026] In this embodiment, the present invention targets the finite element single-physics simulation task to be accelerated. First, the type of the target physical field is identified, including stress-displacement field, seepage field, temperature field, electric field, magnetic field, flow field, etc. Then, based on the governing equations, constitutive relations, geometric equations, and boundary conditions of the target physical field, the physical governing equations of the continuous medium are established and discretized into a finite element numerical format to obtain the equilibrium equations of the single-field finite element system. Based on the equilibrium equations of the finite element system, the physical equation residuals are defined.
[0027] Optionally, the physical equation residuals include one or more combinations of equilibrium equation residuals, constitutive equation residuals, geometric equation residuals, boundary condition residuals, and conservation law residuals, used to characterize the deviation between the single-field simulation results and physical laws, ensuring that the optimization process of the AI model does not violate the physical mechanism, and that optimization is performed only for the single-field solver.
[0028] S2. Construct an AI model, embed the residuals of the above single-field physical equations as constraints into the loss function of the AI model, and simultaneously input the basic parameters of the single-field finite element simulation, outputting a physical constraint loss function to determine the optimal preconditioner type, iterative algorithm, and convergence criterion of the single-field solver.
[0029] In this embodiment, the present invention constructs an AI model for intelligent optimization of solver parameters and selects deep learning neural networks, gradient boosting trees, reinforcement learning networks or hybrid intelligent models; the physical equation residuals obtained in step 1 are embedded into the AI model loss function to generate a physical constraint loss function.
[0030] Optionally, the AI model input layer receives basic parameters for single-field simulation, including material parameters, mesh features, load conditions, boundary conditions, initial state, number of matrix conditions, element type, and working condition type; the AI model output layer outputs the optimal combination of core parameters for the single-field solver, including iterative algorithm type, preconditioner type, convergence criterion type, convergence tolerance, maximum number of iterations, and relaxation factor.
[0031] S3. Construct a training dataset based on the prior physical knowledge of this single field.
[0032] In this embodiment, based on prior physical knowledge and mechanism derivation, the present invention constructs a lightweight training dataset, eliminating the need for large-scale finite element simulation sampling; wherein, the training dataset includes the following dimensional information: 1. Input characteristics: Single-field physical parameters, matrix characteristics, and working condition characteristics under different geotechnical conditions, materials, grids, and loads; 2. Output labels: Optimal solver parameter combinations and convergence performance indices for the corresponding working conditions; 3. Physical constraints: physical equation residuals, conservation law satisfaction, and physical rationality indicators.
[0033] S4. Train the AI model using the training dataset, optimize the model parameters, and ensure that the single-field solver parameters output by the AI model can meet the residual constraints of the single-field physical equation, while achieving the fastest iteration convergence speed and the highest solution efficiency for the single field.
[0034] In this embodiment, the training dataset constructed using S3 is used to train the AI model under the hard constraint of the physical equation residuals. Through algorithms such as backpropagation, gradient descent, and adaptive learning rate optimization, the weight parameters of the AI model are updated so that the solver parameters output by the model can achieve the fastest convergence speed, the fewest iteration steps, the smallest residuals, and the highest stability while satisfying the physical constraints.
[0035] Optionally, during the training process, the present invention also monitors the satisfaction of physical constraints in real time to ensure that the model output does not violate physical laws.
[0036] S5. Embed the trained AI model into the existing finite element solver kernel, perform a single-field simulation of the target, and obtain the optimal solver parameters.
[0037] In this embodiment, the present invention embeds the trained AI model into the existing finite element solver kernel in the form of plug-ins, dynamic link libraries, kernel interfaces, etc., without changing the original simulation process, model construction and mesh generation; for the target single-field simulation task, the AI model automatically reads the input features and infers the optimal solver parameter combination in real time; the finite element solver is automatically configured according to the AI output parameters and starts iterative solution, without the need for manual intervention throughout the process.
[0038] S6. After a single field is solved, the AI model adaptively adjusts the output solver parameters based on the physical equation residuals of the solution results, optimizing the efficiency of subsequent simulations of that single field and forming a closed-loop optimization.
[0039] In this embodiment, after a single field is solved, the present invention automatically extracts information such as the residual of the physical equation, the number of iterations, the convergence time, and the accuracy index of the solved problem, and feeds it back to the AI model. The AI model adaptively updates the weights based on the real-time feedback information and dynamically optimizes the parameter output strategy for subsequent working conditions, forming a closed-loop optimization mechanism of "perception-decision-execution-feedback" to achieve continuous adaptation to complex and variable working conditions and performance improvement.
[0040] like Figure 2As shown, this invention also proposes an intelligent acceleration system for an AI finite element solver based on physical priors, which includes: a hard constraint extraction module, an AI model construction module, a training set construction module, a model parameter optimization module, an optimal parameter acquisition module, and an adaptive adjustment module.
[0041] 1. The hard constraint extraction module is used to determine the physical equation corresponding to a single physical field in the target finite element simulation, and extract the residual of the physical equation as the hard constraint of the AI model.
[0042] In this embodiment, the hard constraint extraction module first identifies the target physical field type for the finite element single-physics simulation task to be accelerated, including stress-displacement field, seepage field, temperature field, electric field, magnetic field, flow field, etc.; then, based on the control equation, constitutive relation, geometric equation, and boundary conditions of the target physical field, it establishes the physical control equation of the continuous medium and discretizes it into a finite element numerical format to obtain the equilibrium equation of the single-field finite element system, and defines the physical equation residuals based on the equilibrium equation of the finite element system.
[0043] Optionally, the physical equation residuals include one or more combinations of equilibrium equation residuals, constitutive equation residuals, geometric equation residuals, boundary condition residuals, and conservation law residuals, used to characterize the deviation between the single-field simulation results and physical laws, ensuring that the optimization process of the AI model does not violate the physical mechanism, and that optimization is performed only for the single-field solver.
[0044] 2. The AI model building module is used to build the AI model. The residuals of the above single-field physical equations are embedded as constraints into the loss function of the AI model. At the same time, the basic parameters of the single-field finite element simulation are input, and the output is the optimal preconditioner type, iterative algorithm and convergence criterion of the single-field solver.
[0045] In this embodiment, the AI model building module constructs an AI model for intelligent optimization of solver parameters, and selects deep learning neural networks, gradient boosting trees, reinforcement learning networks or hybrid intelligent models; the physical equation residuals obtained in step 1 are embedded into the AI model loss function to generate a physical constraint loss function.
[0046] Optionally, the AI model input layer receives basic parameters for single-field simulation, including material parameters, mesh features, load conditions, boundary conditions, initial state, number of matrix conditions, element type, and working condition type; the AI model output layer outputs the optimal combination of core parameters for the single-field solver, including iterative algorithm type, preconditioner type, convergence criterion type, convergence tolerance, maximum number of iterations, and relaxation factor.
[0047] 3. The training set construction module is used to construct a training dataset based on the prior physical knowledge of this single field.
[0048] In this embodiment, based on prior physical knowledge and mechanism derivation, the present invention constructs a lightweight training dataset, eliminating the need for large-scale finite element simulation sampling; wherein, the training dataset includes the following dimensional information: (1) Input characteristics: single-field physical parameters, matrix characteristics, and working condition characteristics under different geotechnical conditions, different materials, different grids, and different loads; (2) Output labels: the optimal solver parameter combination and convergence performance index under the corresponding working conditions; (3) Physical constraints: physical equation residuals, conservation law satisfaction, and physical rationality index.
[0049] 4. The model parameter optimization module is used to train the AI model using the training dataset, optimize the model parameters, and ensure that the single-field solver parameters output by the AI model can meet the residual constraints of the single-field physical equation, while achieving the fastest iteration convergence speed and the highest solution efficiency for the single field.
[0050] In this embodiment, the model parameter optimization module uses the training dataset constructed in step S3 to train the AI model under the hard constraint of the physical equation residuals. Through algorithms such as backpropagation, gradient descent, and adaptive learning rate optimization, the AI model weight parameters are updated so that the solver parameters output by the model achieve the fastest convergence speed, the fewest iteration steps, the smallest residuals, and the highest stability while satisfying the physical constraints.
[0051] Optionally, during the training process, the present invention also monitors the satisfaction of physical constraints in real time to ensure that the model output does not violate physical laws.
[0052] 5. The optimal parameter acquisition module is used to embed the trained AI model into the existing finite element solver kernel and obtain the optimal solver parameters for a single-field simulation of the target.
[0053] In this embodiment, the optimal parameter acquisition module embeds the trained AI model into the existing finite element solver kernel in the form of plug-ins, dynamic link libraries, kernel interfaces, etc., without changing the original simulation process, model construction and mesh generation. For the target single-field simulation task, the AI model automatically reads the input features and infers and outputs the optimal solver parameter combination in real time. The finite element solver is automatically configured according to the AI output parameters and starts iterative solution without manual intervention.
[0054] 6. The adaptive adjustment module is used to adaptively adjust the output solver parameters based on the physical equation residuals of the solution results after a single field is solved, thereby optimizing the efficiency of subsequent simulations of that single field and forming a closed-loop optimization.
[0055] In this embodiment, after a single-field solution is completed, the adaptive adjustment module automatically extracts information such as the physical equation residuals, iteration steps, convergence time, and accuracy indicators from the solution, and feeds this information back to the AI model. The AI model adaptively updates the weights based on the real-time feedback information, dynamically optimizing the parameter output strategy for subsequent operating conditions, forming a closed-loop optimization mechanism of "perception-decision-execution-feedback," achieving continuous adaptation and performance improvement for complex and variable operating conditions. Based on the above disclosure, the present invention also provides an electronic device. The electronic device of this embodiment includes at least one processor and at least one storage medium electrically connected to the processor. The storage medium is electrically connected to the processor, wherein the storage medium stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.
[0056] Based on the same inventive concept, the present invention also provides a storage medium storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described above.
[0057] The foregoing description and accompanying drawings fully illustrate embodiments of the invention to enable those skilled in the art to practice them. Other embodiments may include structural and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Some portions and features of some embodiments may be included or substituted for portions and features of other embodiments. Embodiments of the invention are not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes may be made without departing from their scope. The scope of the invention is limited only by the appended claims.
Claims
1. A method for intelligent acceleration of AI finite element solvers based on physical priors, characterized in that, The method includes, Determine the physical equations corresponding to a single physical field in the target finite element simulation, and extract the physical equation residuals as hard constraints for the AI model based on the physical equations corresponding to a single physical field. Construct an AI model and embed the physical equation residuals as constraints into the loss function of the AI model to obtain a physical constraint loss function; A training dataset is constructed based on this single-field physics prior knowledge; The AI model is trained using the training dataset, and the model parameters are optimized to ensure that the single-field solver parameters output by the AI model can satisfy the residual constraints of the single-field physical equation. The trained AI model is embedded into the existing finite element solver kernel, and the optimal solver parameters are obtained for single-field simulation of the target. After a single-field solution is completed, the AI model adaptively adjusts the output solver parameters based on the physical equation residuals of the solution results, optimizing the efficiency of subsequent simulations of that single field and forming a closed-loop optimization.
2. The intelligent acceleration method for AI finite element solvers based on physical priors as described in claim 1, characterized in that, The process involves determining the physical equations corresponding to individual physical fields in the target finite element simulation, and extracting the residuals of these equations as hard constraints for the AI model. Specifically, this includes... Define the target physical field type; wherein, the target physical field type includes stress-displacement field, seepage field, temperature field, electric field, magnetic field and flow field; Based on the governing equations, constitutive relations, geometric equations and boundary conditions of the target physical field, the physical governing equations of the continuum are established and discretized into a finite element numerical scheme to obtain the equilibrium equations of a single-field finite element system. Based on the equilibrium equations of the finite element system, the physical equation residuals are defined.
3. The intelligent acceleration method for AI finite element solvers based on physical priors as described in claim 2, characterized in that, The physical equation residuals include equilibrium equation residuals, constitutive equation residuals, geometric equation residuals, boundary condition residuals, and conservation law residuals.
4. The intelligent acceleration method for AI finite element solvers based on physical priors as described in claim 1, characterized in that, The training dataset includes input features, output labels, and physical constraints; The input features are single-field physical parameters, matrix features, and working condition features under different geotechnical conditions, different materials, different grids, and different loads. The output labels are the optimal solver parameter combinations and convergence performance indices for the corresponding working conditions. The physical constraints are the physical equation residuals, the degree of satisfaction of conservation laws, and physical rationality indicators.
5. The intelligent acceleration method for AI finite element solvers based on physical priors as described in claim 1, characterized in that, The optimization of the efficiency of subsequent simulations in this single field, forming a closed-loop optimization, specifically includes: Extract the residuals, iteration steps, convergence time, and accuracy metrics of the currently solved physical equations and feed them back to the AI model; The AI model adaptively updates weights based on real-time feedback information, dynamically optimizes parameter output strategies for subsequent operating conditions, and forms a closed-loop optimization mechanism of "perception-decision-execution-feedback".
6. An intelligent acceleration system for an AI finite element solver based on physical priors, characterized in that, The system includes: a hard constraint extraction module, an AI model construction module, a training set construction module, a model parameter optimization module, an optimal parameter acquisition module, and an adaptive adjustment module; The hard constraint extraction module is used to determine the physical equation corresponding to a single physical field in the target finite element simulation, and extract the physical equation residual as the hard constraint condition of the AI model based on the physical equation corresponding to the single physical field. The AI model building module is used to build an AI model, embedding the physical equation residuals as constraints into the loss function of the AI model to obtain a physical constraint loss function. The training set construction module is used to construct a training dataset based on the prior physical knowledge of this single field. The model parameter optimization module is used to train the AI model using the training dataset, optimize the model parameters, and ensure that the single-field solver parameters output by the AI model can meet the residual constraints of the single-field physical equation. The optimal parameter acquisition module is used to embed the trained AI model into the existing finite element solver kernel and obtain the optimal solver parameters for the target single-field simulation. The adaptive adjustment module is used to adaptively adjust the output solver parameters of the AI model based on the physical equation residuals of the solution results after a single field is solved, thereby optimizing the efficiency of subsequent simulations of that single field and forming a closed-loop optimization.
7. The intelligent acceleration system for AI finite element solver based on physical priors as described in claim 6, characterized in that, The hard constraint extraction module is specifically used for, Define the target physical field type; wherein, the target physical field type includes stress-displacement field, seepage field, temperature field, electric field, magnetic field and flow field; Based on the governing equations, constitutive relations, geometric equations and boundary conditions of the target physical field, the physical governing equations of the continuum are established and discretized into a finite element numerical scheme to obtain the equilibrium equations of a single-field finite element system. Based on the equilibrium equations of the finite element system, the physical equation residuals are defined.
8. The intelligent acceleration system for AI finite element solver based on physical priors as described in claim 7, characterized in that, The physical equation residuals include equilibrium equation residuals, constitutive equation residuals, geometric equation residuals, boundary condition residuals, and conservation law residuals.
9. The intelligent acceleration system for AI finite element solver based on physical priors as described in claim 6, characterized in that, The training dataset includes input features, output labels, and physical constraints; The input features are single-field physical parameters, matrix features, and working condition features under different geotechnical conditions, different materials, different grids, and different loads. The output labels are the optimal solver parameter combinations and convergence performance indices for the corresponding working conditions. The physical constraints are the physical equation residuals, the degree of satisfaction of conservation laws, and physical rationality indicators.
10. The intelligent acceleration system for AI finite element solver based on physical priors as described in claim 6, characterized in that, The adaptive adjustment module is specifically used for, Extract the residuals, iteration steps, convergence time, and accuracy metrics of the currently solved physical equations and feed them back to the AI model; The AI model adaptively updates weights based on real-time feedback information, dynamically optimizes parameter output strategies for subsequent operating conditions, and forms a closed-loop optimization mechanism of "perception-decision-execution-feedback".
11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the intelligent acceleration method for the AI finite element solver based on physical priors as described in any one of claims 1-5.
12. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of the intelligent acceleration method for the AI finite element solver based on physical priors as described in any one of claims 1-5.