Digital analogue simulation method

Abstract

The invention relates to the field, and discloses a digital analogue simulation method, which comprises the following steps of: S1, decomposing a control equation of a simultaneous thermal field, an electric field and a force field through a Jacobian matrix, constructing a tight coupling solver, performing implicit solution by adopting a Newton-Rafson iteration method, and adaptively adjusting a time step length according to a residual change rate; s2, automatically comparing temperature field and stress field data output by simulation with experimental measurement point data by using an agent model based on a convolutional neural network and a long-short-term memory network, and adjusting material attribute parameters and restarting simulation by using a gradient descent method when an error exceeds a preset threshold value; and S3, simulation tasks are dynamically allocated to heterogeneous computing resources based on a reinforcement learning strategy, and the strategy optimizes task splitting granularity and node selection by defining a state space, an action space and a reward function. Compared with the prior art, the system has the advantages of higher precision, lower cost and higher intelligence.

Description

Technical Field

[0001] The present invention relates to the technical field of digital simulation technology, and in particular to a digital simulation method. Background Art

[0002] Traditional digital simulation technology has significant shortcomings in multi-physics field coupling, model verification, and computing resource scheduling:

[0003] (1) Low efficiency of multi-physics field coupling: Existing methods often use sequential solutions or weak coupling strategies, resulting in long calculation times and insufficient accuracy. For example, in the thermal-electrical-mechanical coupled simulation of power battery packs, the traditional method has an error of more than 15%, and a single simulation takes more than 24 hours.

[0004] (2) Model verification relies on manual intervention: simulation parameters need to be manually corrected based on experimental data, which is inefficient and cannot provide real-time feedback.

[0005] (3) Rigid resource allocation: Simulation tasks are usually fixedly assigned to a single computing node, making it difficult to dynamically optimize computing resources based on task priority and cost. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the above technical difficulties and provide a digital simulation method with higher precision, lower cost and greater intelligence.

[0007] To solve the above technical problems, the present invention provides a technical solution: a digital simulation method, comprising the following steps:

[0008] S1: The governing equations of the thermal, electric, and force fields are solved simultaneously through Jacobian matrix decomposition, a tightly coupled solver is constructed, and the Newton-Raphson iteration method is used for implicit solution. The time step is adaptively adjusted according to the rate of change of the residual.

[0009] S2: Using a proxy model based on a convolutional neural network (CNN) and a long short-term memory network (LSTM), the temperature and stress field data output by the simulation are automatically compared with the experimental measurement point data. When the error exceeds a preset threshold, the material property parameters are adjusted using the gradient descent method and the simulation is restarted.

[0010] S3: Dynamically allocate simulation tasks to heterogeneous computing resources based on a reinforcement learning strategy. The strategy optimizes task splitting granularity and node selection by defining the state space, action space, and reward function to minimize computational cost and the risk of task timeout.

[0011] As an improvement, the control equations of the thermal field, electric field and force field include:

[0012] Heat conduction equation:

[0013]

[0014] Electrochemical equation (Butler-Volmer equation):

[0015]

[0016] Structural mechanics equation (Hooke's law):

[0017] σ=E ∈

[0018] The equation is reorganized into a unified matrix form through the Jacobian matrix:

[0019] JΔU=-R

[0020] As an improvement, the construction of the proxy model includes the following steps:

[0021] S21: Input data is the grid node temperature and stress data output by simulation;

[0022] S22: Output data is the predicted physical quantity of the experimental measurement point;

[0023] The S23 network structure uses CNN to extract spatial distribution features, LSTM to capture time series dependencies, and the loss function is mean square error (MSE);

[0024] S24: Initialize model weights through the transfer learning framework and fine-tune network parameters using historical experimental data.

[0025] As an improvement, the reinforcement learning strategy specifically includes:

[0026] S31: The state space is defined as the computational capacity (FLOPs), memory requirements, deadlines of the current task queue, as well as the real-time load of each computing node and the unit price of cloud service billing;

[0027] S32: The action space includes the task splitting granularity (number of MPI processes) and node type selection (local GPU, cloud FPGA or CPU);

[0028] S33: The reward function is:

[0029] R=-(∑(t i ×c i )+λ·max(0,t total -t deadline ))

[0030] Among them, t i is the running time of the task at node i, c i is the resource unit price, and λ is the timeout penalty coefficient.

[0031] As an improvement, the implementation of the dynamic resource allocation includes: allocating computationally intensive tasks (such as turbulence simulation) to high-computing-power nodes (FPGA clusters); allocating post-processing tasks to low-cost CPU nodes; and adjusting resource allocation strategies in real time according to task priorities.

[0032] As an improvement, the adjustment of the material property parameters includes thermal conductivity k, Young's modulus E and electrochemical reaction activation energy Ea, and the corrected parameters are directly fed back to the tightly coupled solver to restart the simulation.

[0033] A digital simulation system, characterized by comprising:

[0034] (1) A multi-physics tightly coupled solution module, used to solve the governing equations simultaneously and perform implicit iterative solutions;

[0035] (2) Intelligent verification module, integrating CNN-LSTM proxy model and parameter automatic correction unit;

[0036] (3) Distributed resource scheduling module, equipped with a reinforcement learning engine, supports dynamic task allocation on hybrid cloud platforms (CPU / GPU / FPGA).

[0037] As an improvement,

[0038] The system further includes a visualization interface for displaying simulation results, resource scheduling status, and agent model error analysis reports in real time.

[0039] The advantages of the present invention compared with the prior art are:

[0040] (1) Tightly coupled multi-physics field solution: The thermal, electrical, and force field control equations are solved simultaneously through Jacobian matrix decomposition, and the Newton-Raphson iteration method is used to implicitly solve the equations and adaptively adjust the time step.

[0041] (2) Intelligent verification and parameter correction: Build a proxy model based on CNN-LSTM, automatically compare simulation and experimental data, and correct material property parameters through gradient descent method.

[0042] (3) Dynamic resource scheduling: Based on reinforcement learning strategies, it optimizes task splitting granularity and heterogeneous computing node allocation in real time to minimize computing costs and timeout risks.

[0043] (4) The multi-field coupling algorithm reduces the comprehensive simulation error from 15% to below 8%;

[0044] (5) The resource scheduling engine reduces computing costs by 30% to 50% and shortens task completion time by 40%;

[0045] (6) The intelligent verification module reduces manual intervention by 90% and supports unattended simulation. DETAILED DESCRIPTION

[0046] The embodiments described in this application are only some of the embodiments of the present invention, rather than all of the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work are within the scope of protection of the present invention.

[0047] A digital simulation method comprises the following steps:

[0048] S1: The governing equations of the thermal, electric, and force fields are solved simultaneously through Jacobian matrix decomposition, a tightly coupled solver is constructed, and the Newton-Raphson iteration method is used for implicit solution. The time step is adaptively adjusted according to the rate of change of the residual.

[0049] The governing equations of the thermal field, electric field and force field include:

[0050] Heat conduction equation:

[0051]

[0052] Electrochemical equation (Butler-Volmer equation):

[0053]

[0054] Structural mechanics equation (Hooke's law):

[0055] σ=E ∈

[0056] The equation is reorganized into a unified matrix form through the Jacobian matrix:

[0057] JΔU=-R

[0058] J is the Jacobian matrix, ΔU is the variable increment, and R is the residual vector. The Newton-Raphson iteration method is used to solve it. When the residual norm is less than 10 -6 The time step is adaptively adjusted in the range of 0.1 to 10 ms according to the residual change rate.

[0059] S2: Using a proxy model based on a convolutional neural network (CNN) and a long short-term memory network (LSTM), the temperature and stress field data output by the simulation are automatically compared with the experimental measurement point data. When the error exceeds a preset threshold, the material property parameters are adjusted using the gradient descent method and the simulation is restarted.

[0060] The construction of the proxy model includes the following steps:

[0061] S21: Input data is the grid node temperature and stress data output by simulation;

[0062] S22: Output data is the predicted physical quantity of the experimental measurement point;

[0063] The S23 network structure uses CNN to extract spatial distribution features, LSTM to capture time series dependencies, and the loss function is mean square error (MSE);

[0064] S24: Initialize model weights through the transfer learning framework and fine-tune network parameters using historical experimental data.

[0065] S3: Dynamically allocate simulation tasks to heterogeneous computing resources based on a reinforcement learning strategy. The strategy optimizes task splitting granularity and node selection by defining the state space, action space, and reward function to minimize computational cost and the risk of task timeout.

[0066] The reinforcement learning strategy specifically includes:

[0067] S31: The state space is defined as the computational capacity (FLOPs), memory requirements, deadlines of the current task queue, as well as the real-time load of each computing node and the unit price of cloud service billing;

[0068] S32: The action space includes the task splitting granularity (number of MPI processes) and node type selection (local GPU, cloud FPGA or CPU);

[0069] S33: The reward function is:

[0070] R=-(∑(t i ×c i )+λ·max(0,t total -t deadline ))

[0071] Among them, t i is the running time of the task at node i, c i is the resource unit price, and λ is the timeout penalty coefficient.

[0072] The implementation of the dynamic resource allocation includes: allocating computationally intensive tasks (such as turbulence simulation) to high-computing power nodes (FPGA clusters); allocating post-processing tasks to low-cost CPU nodes; and adjusting resource allocation strategies in real time according to task priorities.

[0073] The material property parameters adjusted include thermal conductivity k, Young's modulus E, and electrochemical reaction activation energy Ea. The corrected parameters are directly fed back to the tightly coupled solver to restart the simulation.

[0074] A digital simulation system, characterized by comprising:

[0075] (1) A multi-physics tightly coupled solution module, used to solve the governing equations simultaneously and perform implicit iterative solutions;

[0076] (2) Intelligent verification module, integrating CNN-LSTM proxy model and parameter automatic correction unit;

[0077] (3) Distributed resource scheduling module, equipped with a reinforcement learning engine, supports dynamic task allocation on hybrid cloud platforms (CPU / GPU / FPGA).

[0078] The system further includes a visualization interface for displaying simulation results, resource scheduling status, and agent model error analysis reports in real time.

[0079] 1. Agent Model Construction

[0080] Input data: mesh node temperature and stress data output by simulation (dimensions: N×M×T, where N is the number of nodes and T is the time step);

[0081] Output data: temperature and strain prediction values ​​at experimental measurement points (such as thermocouple locations);

[0082] Network structure:

[0083] CNN module: 3 convolutional layers to extract spatial features (convolution kernel size 3×3, stride 1);

[0084] LSTM module: 2 layers of LSTM units to capture time series dependencies (hidden layer dimension 128);

[0085] The loss function uses mean square error (MSE) and the optimizer is Adam.

[0086] 2. Automatic parameter correction

[0087] When the error between the surrogate model prediction value and the experimental data is greater than 5%, the following process is triggered:

[0088] (1) Back propagation calculation of the gradient of material parameters (such as thermal conductivity k and Young's modulus E);

[0089] (2) Restart the simulation after adjusting the parameters and update the proxy model weights.

[0090] 3. Application Case: In battery thermal runaway simulation, the proxy model corrects the activation energy parameter Ea based on SEI film decomposition experimental data, reducing the heat generation rate prediction error from 12% to 4%.

[0091] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," 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 explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0092] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

[0093] The above description of the present invention and its embodiments is non-limiting and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by the above description and, without departing from the purpose of the present invention, designs a structure and embodiment similar to the technical solution without creatively designing, they shall fall within the scope of protection of the present invention.

Claims

1. A digital simulation method, characterized in that: The following steps are involved: S1: The governing equations of the thermal, electric, and force fields are solved simultaneously through Jacobian matrix decomposition, a tightly coupled solver is constructed, and the Newton-Raphson iteration method is used for implicit solution. The time step is adaptively adjusted according to the rate of change of the residual. S2: Using a proxy model based on a convolutional neural network (CNN) and a long short-term memory network (LSTM), the temperature and stress field data output by the simulation are automatically compared with the experimental measurement point data. When the error exceeds a preset threshold, the material property parameters are adjusted using the gradient descent method and the simulation is restarted. S3: Dynamically allocate simulation tasks to heterogeneous computing resources based on a reinforcement learning strategy. The strategy optimizes task splitting granularity and node selection by defining the state space, action space, and reward function to minimize computational cost and the risk of task timeout.

2. A digital simulation method according to claim 1, characterized in that: The control equations of the thermal field, electric field and force field include: Heat conduction equation: Electrochemical equation (Butler-Volmer equation): Structural mechanics equation (Hooke's law): σ=E ∈ The equation is reorganized into a unified matrix form through the Jacobian matrix: JΔU=-R 3. A digital simulation method according to claim 1, characterized in that: The construction of the proxy model includes the following steps: S21: Input data is the grid node temperature and stress data output by simulation; S22: Output data is the predicted physical quantity of the experimental measurement point; The S23 network structure uses CNN to extract spatial distribution features, LSTM to capture time series dependencies, and the loss function is mean square error (MSE); S24: Initialize model weights through the transfer learning framework and fine-tune network parameters using historical experimental data.

4. A digital simulation method according to claim 1, characterized in that: The reinforcement learning strategy specifically includes: S31: The state space is defined as the computational capacity (FLOPs), memory requirements, deadlines of the current task queue, as well as the real-time load of each computing node and the unit price of cloud service billing; S32: The action space includes the task splitting granularity (number of MPI processes) and node type selection (local GPU, cloud FPGA or CPU); S33: The reward function is: R=-(∑(t i ×c i )+λ·max(0,t total -t deadline )) Among them, t i is the running time of the task at node i, c i is the resource unit price, and λ is the timeout penalty coefficient.

5. A digital simulation method according to claim 1, characterized in that: The implementation of the dynamic resource allocation includes: allocating computationally intensive tasks (such as turbulence simulation) to high-computing power nodes (FPGA clusters); allocating post-processing tasks to low-cost CPU nodes; and adjusting resource allocation strategies in real time according to task priorities.

6. A digital simulation method according to claim 1, characterized in that: The material property parameters adjusted include thermal conductivity k, Young's modulus E, and electrochemical reaction activation energy Ea. The corrected parameters are directly fed back to the tightly coupled solver to restart the simulation.

7. A digital simulation system according to any one of claims 1 to 6, characterized in that: include: (1) A multi-physics tightly coupled solution module, used to solve the governing equations simultaneously and perform implicit iterative solutions; (2) Intelligent verification module, integrating CNN-LSTM proxy model and parameter automatic correction unit; (3) Distributed resource scheduling module, equipped with a reinforcement learning engine, supports dynamic task allocation on hybrid cloud platforms (CPU / GPU / FPGA).

8. A digital simulation system according to any one of claim 7, characterized in that: The system further includes a visualization interface for displaying simulation results, resource scheduling status, and agent model error analysis reports in real time.

Images