Fluid simulation acceleration system based on AI model

Abstract

The invention discloses a fluid simulation acceleration system and method based on an AI model, and relates to the field of fluid simulation technology and artificial intelligence. According to the system, a fluid feature extraction module, an AI prediction model training module, a simulation data adaptation module and an acceleration simulation execution module are constructed, and a high-complexity calculation process in fluid simulation is predicted and replaced by using an AI model; the technical problems of low calculation efficiency, long time consumption, high hardware resource consumption and the like caused by the fact that traditional fluid simulation depends on a numerical solution are solved. According to the method, the inherent law of fluid motion is learned through the AI model, rapid prediction of the simulation result is achieved, the fluid simulation efficiency is improved by 50%-90% on the premise that it is guaranteed that the simulation precision meets the engineering requirement, and the method is suitable for various scenes needing fluid simulation such as aerospace, automobile engineering and water conservancy engineering and has wide application prospects.

Description

Technical Field

[0001] This invention relates to the intersection of fluid simulation technology and artificial intelligence, specifically to a fluid simulation acceleration system and method based on an AI model. Background Technology

[0002] Fluid simulation technology is widely used in various industrial fields such as aerospace, automotive design, water conservancy engineering, and energy and power. Its core is to solve the Navier-Stokes equations through numerical methods (such as the finite volume method and the finite element method) to simulate the motion state of fluids and their interaction with solid boundaries. However, traditional fluid simulation technology has significant technical bottlenecks: on the one hand, to ensure simulation accuracy, fine meshing of the computational domain is required, leading to an exponential increase in computational load. The simulation process often takes several hours or even days, seriously affecting product development cycles. On the other hand, fluid motion in complex scenarios (such as turbulence and multiphase flow) exhibits strong nonlinearity and temporal correlation. Numerical solutions require a large number of iterative calculations to converge, placing extremely high demands on hardware computing resources (CPU, GPU) and increasing simulation costs.

[0003] In existing technologies, to improve the efficiency of fluid simulation, methods such as simplifying physical models, reducing mesh precision, or optimizing numerical algorithms are commonly used. However, these methods often sacrifice simulation accuracy, making it difficult to meet the needs of high-precision engineering simulations. In recent years, the rapid development of artificial intelligence (AI) technology has provided a new approach to resolving this contradiction. By learning the intrinsic laws of fluid motion through AI models, rapid prediction of simulation results can be achieved, potentially significantly improving simulation efficiency while maintaining accuracy. However, existing AI-based fluid simulation methods still suffer from the following problems: First, AI models lack adaptability to fluid physical properties, making it difficult to meet engineering requirements in terms of prediction accuracy; second, the models have poor generalization ability, only applicable to fluid simulations in specific scenarios; and third, a complete "AI prediction - numerical correction" closed loop has not been formed, making it difficult to cope with simulation needs under complex working conditions. Therefore, there is an urgent need for an AI-based fluid simulation acceleration system and method that balances accuracy and efficiency with strong generalization ability. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fluid simulation acceleration system and method based on AI models. By constructing an AI prediction model that integrates physical characteristics and combining a hybrid simulation mode of "AI rapid prediction + numerical accurate correction", the efficiency of fluid simulation is greatly improved while ensuring simulation accuracy, thus solving the technical problems of long time consumption and high resource consumption in traditional fluid simulation.

[0005] The technical solution adopted in this invention is: A fluid simulation acceleration system based on an AI model, comprising: Fluid feature extraction module: used to acquire initial parameter data for fluid simulation, preprocess the initial parameter data, and extract key feature parameters of fluid motion, including fluid density, viscosity, velocity distribution, boundary condition parameters, and geometric structure parameters; AI prediction model training module: A training dataset is constructed based on historical fluid simulation data, which includes historical initial parameters, corresponding numerical solution simulation results, and key feature parameters; an AI prediction model is constructed using a deep learning network, and the AI ​​prediction model is trained, validated, and optimized using the training dataset to obtain a trained model that meets the preset accuracy requirements; Simulation data adaptation module: Receives the initial input parameters of the fluid to be simulated, calls the fluid feature extraction module to extract the key features of the fluid to be simulated, and converts the key features into an input format that can be recognized by the AI ​​prediction model; Accelerated Simulation Execution Module: The adapted input data is input into the trained AI prediction model, and the AI ​​prediction model outputs the simulation result prediction value of the fluid to be simulated. If the error of the prediction value meets the preset threshold, the prediction value is directly output as the final simulation result. If the error does not meet the preset threshold, the prediction value is used as the initial iteration value of the numerical solution, the simplified numerical calculation process is started, and the corrected simulation result is output.

[0006] Furthermore, the preprocessing process of the fluid feature extraction module includes: data cleaning to remove outliers and missing values, data normalization to unify parameter magnitudes, and data dimensionality reduction to retain core features. The data dimensionality reduction process uses principal component analysis (PCA) or linear discriminant analysis (LDA) algorithms.

[0007] Furthermore, the AI ​​prediction model adopts a hybrid deep learning architecture, including a feature encoding layer, a temporal prediction layer, and a result decoding layer; the feature encoding layer uses a convolutional neural network (CNN) to extract spatial features, the temporal prediction layer uses a long short-term memory network (LSTM) or a gated recurrent unit (GRU) to capture the temporal correlation of fluid motion, and the result decoding layer uses a fully connected network to output the simulation result prediction value.

[0008] Furthermore, it also includes a model optimization module, which optimizes the AI ​​prediction model through adaptive learning rate adjustment, regularization processing, and ensemble learning strategy. The ensemble learning strategy uses weighted fusion of prediction results from multiple sub-models, with the weights dynamically adjusted based on the validation set accuracy of the sub-models.

[0009] An AI model-based method for accelerating fluid simulation includes the following steps: S1: Data Acquisition and Preprocessing: Collect historical fluid simulation data under different scenarios, including initial parameters, boundary conditions, geometric parameters and corresponding numerical solution simulation results; clean, normalize and reduce the dimensionality of the historical data, extract key feature parameters, and construct training datasets and validation datasets; S2: AI Prediction Model Construction and Training: Based on the physical characteristics of fluid motion, a hybrid deep learning model integrating spatial feature extraction and temporal prediction is constructed; the model is trained using a training dataset, the model accuracy is monitored using a validation dataset, and adaptive optimization algorithms are used to adjust the model parameters until the model accuracy meets the preset requirements; S3: Simulation Data Adaptation: Obtain the parameters of the target object to be simulated, including geometric parameters, fluid physics parameters, and boundary condition parameters; extract key features of the target object through feature extraction algorithms and convert them into data that matches the input format of the AI ​​prediction model; S4: AI-accelerated simulation execution: Input the adapted target object data into the trained AI prediction model to obtain the simulation result prediction value; calculate the error between the prediction value and the numerical solution benchmark value. If the error is ≤ a preset threshold, output the prediction value as the final simulation result; if the error is > a preset threshold, use the prediction value as the initial iteration value of the numerical solution, execute a simplified numerical calculation process (reduce the number of iterations or simplify the computation grid), obtain the corrected simulation result and output it. S5: Model Iteration and Update: Collect new simulation data (including AI prediction results and actual verification results), regularly perform incremental training on the AI ​​prediction model, optimize model parameters, and improve the model's adaptability to complex scenarios.

[0010] Furthermore, the adaptive optimization algorithm in step S2 employs either the Adam algorithm or the RMSProp algorithm, and the evaluation metrics for model accuracy include mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 The default requirement is R. 2 ≥0.92 and RMSE≤5%.

[0011] Furthermore, the simplified numerical calculation process described in step S4 includes: optimizing the computational grid partitioning based on the AI ​​prediction results, using coarse grids for computation in regions with stable fluid motion, and retaining fine grids for computation in regions with high gradients; at the same time, using the AI ​​prediction results as the initial values ​​for the numerical solution to reduce the number of iterations required for convergence.

[0012] Furthermore, the incremental training described in step S5 adopts an online learning mode, which only updates the parameters locally on the new data, avoiding retraining on the full dataset and reducing the time cost of model updates.

[0013] The beneficial effects obtained by this invention are as follows: 1. Significantly improve simulation efficiency: By quickly predicting fluid simulation results through AI models and combining them with simplified numerical correction processes, the fluid simulation time can be shortened by 50%-90% while ensuring accuracy, solving the problem of long simulation time in traditional simulation and significantly improving product development efficiency; 2. Ensure simulation accuracy: Employ a hybrid deep learning model that integrates physical characteristics, combined with a closed-loop approach of "AI prediction + numerical correction," ensuring the simulation results' coefficient of determination R0 is high. 2 ≥0.92, meeting engineering-grade high precision requirements; 3. Strong generalization ability: Through training with historical data from multiple scenarios, combined with ensemble learning and incremental training strategies, the model can adapt to the fluid simulation needs of different fields such as aerospace, automotive engineering, and water conservancy engineering. 4. Reduced resource consumption: The AI ​​prediction process has low hardware resource requirements, and the simplified numerical calculation process reduces the number of iterations and meshes, which greatly reduces the hardware resource consumption and computing costs in the simulation process. Detailed Implementation

[0014] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0015] 1. AI Model-Based Fluid Simulation Acceleration System This system includes a fluid feature extraction module, an AI prediction model training module, a simulation data adaptation module, an accelerated simulation execution module, and a model optimization module. These modules work together to achieve efficient acceleration of fluid simulation.

[0016] Fluid Feature Extraction Module: The core function of this module is to extract key features that characterize the essence of fluid motion from the original simulation data. First, it receives the initial parameter data from the fluid simulation, including fluid physical parameters (density, viscosity, thermal conductivity, etc.), geometric parameters (channel dimensions, boundary shapes, etc.), and boundary condition parameters (inlet velocity, outlet pressure, wall temperature, etc.). Then, the initial data is preprocessed by removing outliers and missing values ​​caused by measurement errors through data cleaning, and using min-max normalization or Z-score normalization to unify the parameter magnitudes, avoiding the impact of parameter magnitude differences on model training performance. Finally, principal component analysis (PCA) or linear discriminant analysis (LDA) algorithms are used to reduce the data dimensionality, eliminate redundant features, retain core feature parameters strongly correlated with the fluid motion results, and reduce the computational complexity of the model.

[0017] AI Prediction Model Training Module: This module employs a hybrid deep learning architecture, fully combining the spatial feature extraction capabilities of Convolutional Neural Networks (CNNs) and the temporal prediction capabilities of Long Short-Term Memory Networks (LSTMs). The feature encoding layer uses CNN kernels to convolve high-dimensional feature data, extracting spatial distribution features of fluid motion (such as velocity gradients and pressure field distributions). The temporal prediction layer uses LSTM or GRU networks to capture the temporal correlations of fluid motion, suitable for non-steady-state fluid simulation scenarios. The result decoding layer uses a fully connected network to map the extracted features to the final simulation results (such as velocity distribution, pressure values, and fluid forces). During model training, a training dataset is constructed based on historical fluid simulation data. The Adam or RMSProp algorithm is used to adaptively adjust the learning rate, and L2 regularization is used to prevent overfitting. Simultaneously, an ensemble learning strategy is introduced, constructing multiple sub-models and dynamically adjusting weights based on validation set accuracy to improve the model's generalization ability and prediction stability.

[0018] Simulation Data Adaptation Module: This module is responsible for converting the raw parameters of the fluid to be simulated into an input format recognizable by the AI ​​prediction model. First, it receives the target object parameters input by the user, calls the feature extraction algorithm from the fluid feature extraction module to extract the key features of the target object, and then performs format conversion and dimension matching on the feature data according to the input requirements of the AI ​​prediction model. This ensures that the data can be directly input into the model for prediction calculations, reducing information loss during the data conversion process.

[0019] Accelerated Simulation Execution Module: This module is the core execution unit for accelerating simulations, employing a hybrid approach of "AI prediction + numerical correction." First, the adapted simulation data is input into the trained AI prediction model, which quickly outputs the predicted simulation results. This process takes only 1%-10% of the time of traditional numerical methods. Then, the error between the predicted value and the baseline value of the numerical solution is calculated (using MAE, RMSE, and R-squared values). 2 As an evaluation indicator, if the error meets the preset threshold (such as R), 2 If the error is greater than or equal to 0.92 and RMSE ≤ 5%, the predicted value is directly output as the final simulation result, achieving extreme acceleration. If the error does not meet the threshold, the predicted value is used as the initial iteration value of the numerical solution, and a simplified numerical calculation process is started. The AI ​​prediction results identify the high gradient region and the stable region of fluid motion. Coarse mesh calculation is used for the stable region, and fine mesh calculation is retained for the high gradient region. At the same time, the high approximation of the predicted value is used to reduce the number of numerical iterations to converge, and finally, the corrected high-precision simulation result is output.

[0020] Model Optimization Module: This module is used to continuously improve the performance of AI prediction models. It optimizes model parameters through adaptive learning rate adjustment (dynamically adjusting the learning rate according to the changes in the loss function during training), regularization (avoiding model overfitting), and ensemble learning strategies (weighted fusion of multiple sub-models). It also supports online incremental training. After collecting new simulation data (including AI prediction results and actual verification results), it only updates local parameters of the model without retraining on full data, reducing the time cost of model updates and improving the model's adaptability to new scenarios and working conditions.

[0021] 2. AI Model-Based Accelerated Fluid Simulation Method This method comprises five steps: data acquisition and preprocessing, AI prediction model construction and training, adaptation to simulation data, AI-accelerated simulation execution, and model iterative updates. The specific process is as follows: Step S1: Data Acquisition and Preprocessing: First, collect historical fluid simulation data from different scenarios such as aerospace, automotive engineering, and hydraulic engineering to ensure that the data covers different geometric structures, fluid physical parameters, and boundary conditions, thereby improving the model's generalization ability. Historical data includes initial parameters (fluid density, viscosity, inlet velocity, etc.), boundary conditions (wall constraints, temperature boundaries, etc.), geometric parameters (channel dimensions, obstacle shapes, etc.), and corresponding numerical simulation results (calculated using traditional high-precision numerical methods, serving as benchmark data). Then, preprocess the historical data: remove outliers using the 3σ criterion or box plot method, and fill in missing values ​​using interpolation; normalize all parameters, mapping parameter values ​​to the [0,1] or [-1,1] interval; use PCA or LDA algorithms for data dimensionality reduction, retaining core feature parameters, and construct a training dataset (70%-80%) and a validation dataset (20%-30%).

[0022] Step S2: AI Prediction Model Construction and Training: Based on the physical properties of fluid motion (such as mass conservation, momentum conservation, and energy conservation), a hybrid deep learning model integrating CNN and LSTM is constructed to ensure that the model can capture the spatial features and temporal correlations of fluid motion. The training dataset is input into the model, and the Adam algorithm is used for model training with mean squared error (MSE) as the loss function. During training, the MAE, RMSE, and R-value of the model are calculated every 10 iterations using the validation dataset. 2 Monitor changes in model accuracy; when model accuracy meets preset requirements (R... 2When the accuracy is ≥0.92 and RMSE≤5%, training is stopped, and the trained AI prediction model is obtained. If the accuracy does not meet the standard, the model structure is adjusted (such as increasing the number of convolutional layers or adjusting the number of LSTM units) or the training parameters are optimized (such as learning rate and regularization coefficient), and training is repeated.

[0023] Step S3: Simulation Data Adaptation: The user inputs the parameters of the target object to be simulated, including geometric parameters (such as aircraft wing dimensions, automobile air intake shape, etc.), fluid physics parameters (such as air density, viscosity, etc.), and boundary condition parameters (such as flight speed, air intake inlet pressure, etc.). The feature extraction algorithm of the fluid feature extraction module is called to preprocess and extract features from the target object parameters to obtain key feature parameters; then, the key feature parameters are converted into data that matches the input format of the AI ​​prediction model (such as adjusting data dimensions and unifying data types) to ensure that the data can be directly input into the model for prediction.

[0024] Step S4: AI-Accelerated Simulation Execution: Input the adapted target object data into the trained AI prediction model. The model quickly outputs the predicted simulation results (such as pressure distribution on the wing surface, velocity distribution in the air intake, etc.). This process typically takes only 1%-10% of the time of traditional numerical methods. Calculate the error between the predicted value and the baseline value of the numerical method. If the error meets a preset threshold, the predicted value is directly output as the final simulation result, achieving rapid simulation. If the error does not meet the threshold, the predicted value is used as the initial iteration value for the numerical method (such as the finite volume method), initiating a simplified numerical calculation process: Optimize the computational grid based on the AI ​​prediction results, reducing the number of grids in stable regions and retaining fine grids in high-gradient regions; simultaneously, utilize the high approximation of the predicted values ​​to reduce the number of numerical iterations required for convergence (traditional numerical methods require hundreds of iterations, while this method only requires dozens), finally obtaining and outputting the corrected high-precision simulation results.

[0025] Step S5: Model Iterative Update: Collect new simulation data, including AI prediction results, numerical correction results, and actual engineering verification results (such as experimental measurement data), add them to the dataset, and periodically perform incremental training on the AI ​​prediction model. Incremental training adopts an online learning mode, updating only local parameters of the model to avoid retraining on the entire dataset and reduce time costs; through continuous iterative updates, improve the model's adaptability to complex scenarios and new working conditions, and ensure continuous optimization of model performance.

[0026] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fluid simulation acceleration system based on an AI model, characterized in that, include: Fluid feature extraction module: used to acquire initial parameter data for fluid simulation, preprocess the initial parameter data, and extract key feature parameters of fluid motion, including fluid density, viscosity, velocity distribution, boundary condition parameters, and geometric structure parameters; AI prediction model training module: Constructs a training dataset based on historical fluid simulation data, which includes historical initial parameters, corresponding numerical solution simulation results, and key feature parameters; A deep learning network is used to build an AI prediction model. The AI ​​prediction model is trained, validated and optimized using a training dataset to obtain a trained model that meets the preset accuracy requirements. Simulation data adaptation module: Receives the initial input parameters of the fluid to be simulated, calls the fluid feature extraction module to extract the key features of the fluid to be simulated, and converts the key features into an input format that can be recognized by the AI ​​prediction model; Accelerated Simulation Execution Module: Inputs the adapted input data into the trained AI prediction model, and outputs the simulation result prediction value of the fluid to be simulated through the AI ​​prediction model; If the error of the predicted value meets the preset threshold, the predicted value will be directly output as the final simulation result. If the error does not meet the preset threshold, the predicted value will be used as the initial iteration value of the numerical solution, the simplified numerical calculation process will be started, and the corrected simulation results will be output.

2. The fluid simulation acceleration system based on an AI model according to claim 1, characterized in that, The preprocessing process of the fluid feature extraction module includes: data cleaning to remove outliers and missing values, data normalization to unify parameter magnitudes, and data dimensionality reduction to retain core features. The data dimensionality reduction process uses principal component analysis (PCA) or linear discriminant analysis (LDA) algorithms.

3. The fluid simulation acceleration system based on an AI model according to claim 1, characterized in that, The AI ​​prediction model employs a hybrid deep learning architecture. It includes a feature encoding layer, a temporal prediction layer, and a result decoding layer; the feature encoding layer uses a convolutional neural network (CNN) to extract spatial features, the temporal prediction layer uses a long short-term memory network (LSTM) or a gated recurrent unit (GRU) to capture the temporal correlation of fluid motion, and the result decoding layer uses a fully connected network to output the simulation result prediction value.

4. The fluid simulation acceleration system based on an AI model according to claim 1, characterized in that, It also includes a model optimization module, which optimizes the AI ​​prediction model through adaptive learning rate adjustment, regularization processing and ensemble learning strategy. The ensemble learning strategy uses weighted fusion of the prediction results of multiple sub-models, and the weights are dynamically adjusted based on the validation set accuracy of the sub-models.

5. A method for accelerating fluid simulation based on AI models, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing: Collect historical fluid simulation data under different scenarios, including initial parameters, boundary conditions, geometric parameters and corresponding numerical solution simulation results; clean, normalize and reduce the dimensionality of the historical data, extract key feature parameters, and construct training datasets and validation datasets; S2: AI Prediction Model Construction and Training: Based on the physical characteristics of fluid motion, a hybrid deep learning model integrating spatial feature extraction and temporal prediction is constructed; the model is trained using a training dataset, the model accuracy is monitored using a validation dataset, and adaptive optimization algorithms are used to adjust the model parameters until the model accuracy meets the preset requirements; S3: Simulation Data Adaptation: Obtain the parameters of the target object to be simulated, including geometric parameters, fluid physics parameters, and boundary condition parameters; extract the key features of the target object through feature extraction algorithms and convert them into data that matches the input format of the AI ​​prediction model; S4: AI-accelerated simulation execution: Input the adapted target object data into the trained AI prediction model to obtain the simulation result prediction value; Calculate the error between the predicted value and the baseline value of the numerical solution. If the error is less than or equal to a preset threshold, output the predicted value as the final simulation result. If the error is greater than the preset threshold, the predicted value will be used as the initial iteration value of the numerical solution, and a simplified numerical calculation process will be executed (reducing the number of iterations or simplifying the computation grid) to obtain and output the corrected simulation results. S5: Model Iteration and Update: Collect new simulation data (including AI prediction results and actual verification results), regularly perform incremental training on the AI ​​prediction model, optimize model parameters, and improve the model's adaptability to complex scenarios.

6. The method for accelerating fluid simulation based on an AI model according to claim 5, characterized in that, The adaptive optimization algorithm described in step S2 employs either the Adam algorithm or the RMSProp algorithm. The evaluation metrics for model accuracy include mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 The default requirement is R. 2 ≥0.92 and RMSE≤5%.

7. The method for accelerating fluid simulation based on an AI model according to claim 5, characterized in that, The simplified numerical calculation process described in step S4 includes: optimizing the computational grid partitioning based on AI prediction results, using coarse grids for computation in regions with stable fluid motion, and retaining fine grids for computation in regions with high gradients; at the same time, using AI prediction results as the initial values ​​for the numerical solution to reduce the number of iterations required for convergence.

8. The method for accelerating fluid simulation based on an AI model according to claim 5, characterized in that, The incremental training described in step S5 adopts an online learning mode, which only updates local parameters on newly added data, avoiding retraining on all data and reducing the time cost of model updates.