Numerical simulation accelerated convergence method based on artificial intelligence prediction
By intelligently interacting the AI prediction network and numerical simulation network based on the U-Net architecture, the solution status is monitored in real time and the flow field parameters are predicted. This solves the problems of high computational cost and long cycle in the numerical simulation of aero-engines, and achieves efficient and accurate simulation acceleration convergence, which is suitable for complex working conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-06
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies for numerical simulation of aero-engines suffer from high computational costs, long computation cycles, high hardware requirements, low accuracy, and low efficiency under complex operating conditions. In particular, communication bottlenecks and unbalanced loads are severe when dealing with large-scale grid computing.
An AI prediction network based on the U-Net architecture is adopted. By constructing an intelligent interaction between the network and the numerical simulation network, the solution status is monitored in real time. At the oscillating convergence node, the solver is paused, and the flow field parameters are predicted by the prediction network to generate update vectors, thus skipping the time-consuming oscillation stage and achieving accelerated convergence.
It significantly shortens the computation time of numerical simulation, saves costs, improves the efficiency and accuracy of simulation, has good versatility, is suitable for complex working conditions such as strongly nonlinear and unsteady flow, and does not depend on a specific hardware architecture.
Smart Images
Figure CN121835304A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aerospace technology, and in particular to a method, system, device and medium for accelerating convergence of numerical simulation based on artificial intelligence prediction. Background Technology
[0002] Numerical simulation is a crucial component of modern fluid mechanics research. By solving partial differential equations using computers, flow field solutions in fluid mechanics can be provided with a certain level of accuracy. In the field of aero-engines, numerical simulation improves research and development efficiency and reduces design costs. Furthermore, almost all research in modern aero-engines relies heavily on the assistance of numerical simulation. To obtain solutions that more closely approximate physics, engineers have improved computational methods and developed new solution methods, enabling numerical simulation to solve computational cases with hundreds of millions of grid points. However, this progress is hampered by the slow development of computer hardware, continuously rising computational costs, and increasingly longer computation cycles.
[0003] In recent years, artificial intelligence (AI) technology has provided a novel solution to overcome this bottleneck. AI methods are no longer limited to optimizing existing physics algorithms, but can learn physical laws from historical simulation data. This technology offers a completely new and feasible approach to accelerating numerical simulation.
[0004] Existing technologies include optimizations to address oscillations in the computation process. These include the golden section technique to introduce loosely coupled computation, guiding the wing structure to monotonically converge from its initial position to its static equilibrium position, thus accelerating computation by eliminating oscillations during convergence. Other methods utilize Gauss-Seidel and Shure completion methods to reduce significant oscillations during computation.
[0005] Meanwhile, traditional parallel computing methods, as the mainstream approach, also deserve attention. Among them, the domain decomposition method divides a complex computational domain into multiple sub-regions, each computed by a different processor. Designing efficient dedicated parallel solvers to address bottlenecks in the solution process can significantly improve parallel efficiency.
[0006] There are also methods for optimizing and accelerating computation for specific hardware. For example, parallel optimization for GPUs is widely used. It accelerates computation by leveraging the GPU's strong parallel computing design through strategies such as overlapping computation and communication and memory access optimization.
[0007] AI-based solver replacement is a completely new approach. By incorporating physical information into constraints through physical information neural networks, it enables the prediction of flow fields, solving some computational problems and greatly reducing computation time.
[0008] The disadvantages of existing technology are:
[0009] 1. While existing technologies have achieved significant results in oscillation optimization methods, their acceleration capabilities are often limited by the inherent defects and problem-dependent nature of the algorithms themselves, and they still fall short when dealing with complex working conditions such as strong nonlinearity and unsteadiness.
[0010] 2. Traditional parallel methods face communication bottlenecks and load imbalances when dealing with a large number of grids, which reduces parallel efficiency. Moreover, larger-scale computing places higher demands on hardware and increases costs.
[0011] 3. Methods that optimize and accelerate specific hardware lack mature computing software and are limited by the memory bandwidth of GPUs, resulting in high costs for large-scale computing.
[0012] 4. The accuracy of the physical information neural network method is usually lower than that of the traditional method, which is not of reference value for modern aero-engine research and is difficult to solve the accuracy problem. Summary of the Invention
[0013] The purpose of this invention is to provide a method, system, device, and medium for accelerating convergence in numerical simulation based on artificial intelligence prediction. This method accelerates computation by skipping the oscillating convergence stage in existing numerical simulation calculations. Compared to existing commercial software acceleration methods, this invention solves the problem of excessively long convergence times while saving computational costs. Furthermore, compared to oscillation optimization methods, this invention can be extended to complex operating conditions, exhibiting strong versatility. This invention uses existing numerical simulation software as a foundation and does not rely on the development of new hardware. The artificial intelligence method employed in this invention does not completely replace numerical simulation calculations, thus maintaining computational accuracy. Moreover, since the final step is still completed by the numerical simulation software, there are no issues with physical constraints.
[0014] This invention provides the following technical solution:
[0015] Firstly, a numerical simulation convergence acceleration method based on artificial intelligence prediction includes the following steps:
[0016] S1. Configure the numerical simulation network and build the connection network.
[0017] The configuration parameters of the numerical simulation network are obtained, including the physical model, boundary conditions, and initial field. The construction of the network includes setting up data extraction channels, data writing channels, and building an update vector generation module to realize intelligent interaction between the user and the numerical simulation network, perform real-time data extraction and writing, generate update vectors, and control the solver of the numerical simulation network.
[0018] S2, Constructing the prediction network
[0019] A U-Net main network, comprising an encoder and a decoder, is constructed. The encoder downsamples the flow field grid parameters from multiple historical time steps to extract spatial features, while the decoder restores spatial resolution and reconstructs the flow field through deconvolution and skip connections. A spatiotemporal attention module is integrated into the U-Net. The spatial attention module focuses on key regions of the flow field, while the temporal attention module learns the dynamic evolution of flow field states across multiple consecutive time steps. The network is trained using numerical simulation data, with flow field data from multiple time steps in the previous stage as input and the corresponding near-converged flow field state as the target. This allows the network to learn the nonlinear mapping relationship from the initial stage of oscillation to the later stage of convergence, thereby outputting a predicted flow field with a complete physical structure, resulting in an optimized prediction network.
[0020] S3. Perform numerical simulations and calculations to accelerate convergence.
[0021] The connection network monitors the solution status of the numerical simulation network in real time, pauses the solver at the residual oscillation node, and extracts multi-step flow field data. The multi-step flow field data is transmitted to the prediction network, which predicts the flow field parameters of the globally approximate convergent solution. The update vector is calculated based on the flow field parameters to obtain the prediction result. Finally, the connection network writes the prediction result back to the numerical simulation network, overwriting the current state of the solver and guiding it to restart the iteration based on a high-quality initial field, thereby accelerating convergence.
[0022] Preferably, constructing the connection network in step S1 includes the following steps:
[0023] S1.1 Constructing data extraction and data writing channels
[0024] Extraction and input functions are implemented using the solver's native interface. The solver is rearranged to meet the parameter classification requirements of different spaces, times, and categories, which facilitates training and prediction.
[0025] S1.2, Constructing the Update Vector Generation Module
[0026] Solvers used for numerical simulations employ various operating methods, including finite difference, finite element, spectral methods, and finite volume methods. These solvers combine the discrete equations of all variables into a globally coupled system of equations, expressed as follows:
[0027] (1)
[0028] In the above formula, [A] is a coefficient matrix. [Φ] is the correction vector, and [R] is the residual vector. After solving for the correction vector, the initial [Φ], which is the variable value at the beginning of this iteration, is updated. All solved variables are updated at the same time. A lightweight neural network is constructed, and by training it to capture the relationship between the coefficient matrix and the correction vector, the accurate estimation of the correction vector is achieved.
[0029] S1.3 Solver Control Interface
[0030] The control interface is implemented through the solver's script control function. The control interface sends "pause" and "continue" commands to the solver to ensure that the calculation is interrupted and restarted.
[0031] Preferably, step S2, constructing the prediction network, includes the following steps:
[0032] S2.1 Network Architecture Setup
[0033] The main network architecture is based on the U-Net artificial intelligence structure. The encoder consists of multiple convolutional and pooling layers, which are responsible for downsampling the current flow field mesh parameters, identifying the main structure, and then extracting abstract representations of these core features. Each downsampling step reduces the spatial resolution but increases the number of feature channels, enhancing the learning accuracy of the flow field features. The decoder gradually restores the spatial resolution through deconvolution operations. Each layer of the decoder receives feature maps from layers of the same scale as the encoder, achieving skip connections and generating prediction results with complete physical structure and data that are closer to the final convergence result.
[0034] The prediction network is input by taking the results of ten iterative calculations before the key nodes. At the same time, a spatiotemporal attention mechanism is introduced. The spatial attention module automatically assigns higher weights to these key regions, making the network more focused during prediction. The temporal attention mechanism is used to learn the evolution law between continuous states and understand the generation, development and dissipation of vortices, or the dynamic processes such as the forward movement and hysteresis of shock waves.
[0035] S2.2, Artificial Intelligence Network Training
[0036] The training data for the prediction network consists of a large number of completed, high-precision numerical simulation examples. For each example, after the entire computation process is completed and convergence is achieved, the rapid descent region during convergence is first determined based on the number of grid cells, and the last ten steps are used as input features. The state of the same example nearing convergence is used as the training target value. In this way, the network learns the nonlinear mapping relationship from the initial stage of computational oscillation to the later stage of computational convergence. To improve the model's generalization ability, the training data is standardized, and various data augmentation techniques are employed, including perturbing the extracted parameters within a reasonable range to indirectly increase the sample size and enhance anti-interference ability, ensuring that the trained model has good adaptability to different operating conditions.
[0037] Preferably, step S3 includes the following steps:
[0038] S3.1 Initialization and Monitoring Phase:
[0039] S3.1.1 First, set the physical model, boundary conditions, and initial field in the numerical simulation network, and then start the calculation;
[0040] S3.1.2 Connect the network to activate and start monitoring the computational status of the numerical simulation network, especially to determine whether the preset convergence oscillation key point has been reached; the data logger is used to record the grid node parameters obtained from the ten calculation steps prior to the current calculation step;
[0041] S3.2, Calculation Interruption and Data Extraction Stage:
[0042] S3.2.1 When the solver of the numerical simulation network reaches a critical node that the program determines is in the early stage of convergence oscillation, specifically at the Nth step when the residual curve begins to enter high-frequency, small-amplitude oscillation, step S3.2.2 is executed; when the solver program determines that it is not a critical node in the early stage of convergence oscillation, it returns to step S3.1.2.
[0043] S3.2.2 The connection network sends a solver calculation pause signal to the solver of the numerical simulation network; the solver maintains the current calculation state to ensure that the physical quantities on all grid cells remain intact; subsequently, the data extraction channel of the connection network starts working, extracting the grid point parameters of all regions in the current iteration step and the ten steps before the current iteration step from the solver, and summarizing them into the data format input to the artificial intelligence prediction network;
[0044] S3.3, Acceleration Stage of Artificial Intelligence Prediction
[0045] S3.3.1 Artificial intelligence prediction network predicts parameters. The grid point parameters extracted in step S3.2.2 are processed by the prediction network. The prediction span of the prediction network depends on the complexity of the grid. The predicted parameters are the parameters of all points in the entire grid.
[0046] S3.3.2 Calculation and generation of update vector: The update vector generation module inside the network calculates the predicted parameters and generates a new update vector. The update vector provides an iterative direction for the iterative calculation to match the computational requirements of the numerical simulation network solver when it restarts from a new state.
[0047] The S3.3.3 parameter write-back numerical simulation solver writes the grid point parameters and update vectors predicted by artificial intelligence into the numerical simulation network solver through the connection network, overwriting the previous state; finally, the connection network sends a calculation restart signal to the solver, and the numerical simulation iterative calculation continues based on the obtained new parameters.
[0048] S3.4 Numerical Simulation Calculation Final Convergence and Output
[0049] In the computation monitor of the numerical simulation network, the user determines whether the required computational accuracy has been achieved. If the required accuracy is achieved, the computation is terminated and the results are output. If the required accuracy is not achieved, the numerical simulation solver continues to compute until the requirements are met.
[0050] Secondly, the present invention provides a numerical simulation accelerated convergence system based on artificial intelligence prediction, comprising:
[0051] The configuration of the numerical simulation network and the construction of the connection network module are used to obtain the configuration parameters of the numerical simulation network, including the physical model, boundary conditions and initial field; the construction of the connection network includes setting up data extraction channels, data writing channels and constructing an update vector generation module to realize intelligent interaction between the user and the numerical simulation network, perform real-time data extraction and writing, generate update vectors and control the solver of the numerical simulation network;
[0052] A prediction network module is constructed to build the main U-Net network, which includes an encoder and a decoder. The encoder downsamples the flow field grid parameters of multiple historical time steps to extract spatial features, and the decoder restores the spatial resolution to reconstruct the flow field through deconvolution and skip connections. A spatiotemporal attention module is integrated into the U-Net. The spatial attention module is used to focus on key regions of the flow field, and the temporal attention module is used to learn the dynamic evolution law between the flow field states of multiple consecutive time steps. The network is trained using numerical simulation data. The flow field data of multiple time steps in the previous stage are used as input, and the corresponding near-converged flow field state is used as the target. The network learns the nonlinear mapping relationship from the initial stage of oscillation to the later stage of convergence, thereby outputting a predicted flow field with a complete physical structure, resulting in an optimized prediction network.
[0053] The numerical simulation and computational acceleration convergence module is used to enable the connection network to monitor the solution status of the numerical simulation network in real time, pause the solver at the residual oscillation node, and extract multi-step flow field data; transmit the multi-step flow field data to the prediction network, predict the flow field parameters of the globally approximate convergent solution through the prediction network, calculate the update vector based on the flow field parameters, and obtain the prediction result; finally, the connection network writes the prediction result back to the numerical simulation network, overwrites the current state of the solver, and guides it to restart the iteration based on a high-quality initial field, thereby achieving accelerated convergence.
[0054] Thirdly, the present invention provides a computer device, the computer device including an input interface and an output interface, and further including: a processor adapted to implement one or more instructions;
[0055] A computer storage medium storing one or more instructions, the one or more instructions being adapted to be loaded and executed by the processor for the numerical simulation accelerated convergence method described above.
[0056] Fourthly, the present invention provides a computer storage medium storing one or more instructions, the one or more instructions being adapted to be loaded and executed by a processor to perform the above-described numerical simulation accelerated convergence method.
[0057] Compared with the prior art, the present invention has the following beneficial effects:
[0058] This invention provides a method, system, device, and medium for accelerating convergence in numerical simulation based on artificial intelligence prediction. By creating a prediction network, this AI-enabled network replaces the convergence process of traditional numerical simulation calculations, significantly accelerating the convergence process. The prediction network precisely skips the most time-consuming oscillating convergence phase in the simulation, directly providing the solver with high-quality initial conditions close to the convergent solution, thereby greatly shortening computation time and saving costs. Simultaneously, since the final convergence result is still achieved by validated traditional numerical simulation software, the accuracy and physical plausibility of the solution are fundamentally guaranteed.
[0059] For the same cascade flow field calculation problem with 3.4 million grids, the convergence time of this invention is reduced by more than 60% compared with the convergence time of traditional CFD calculation, effectively saving computing resources and improving the efficiency of numerical simulation. This method has excellent versatility. The prediction network can effectively capture various complex flow field characteristics through the spatiotemporal attention mechanism, performs well in various working conditions such as strongly nonlinear and unsteady flow, and does not depend on a specific hardware architecture. Attached Figure Description
[0060] Figure 1 This is a flowchart illustrating the numerical simulation acceleration convergence method based on artificial intelligence prediction provided in an embodiment of the present invention. Detailed Implementation
[0061] This invention provides an intelligent method, system, device, and medium for accelerating numerical simulation convergence. The method integrates a U-Net architecture with a spatiotemporal attention mechanism to improve the prediction network. This improved network can predict near-convergence results from current and historical computational states. It also provides methods for real-time data extraction, update vector generation, and intelligent solver control to support the prediction network. Furthermore, the acceleration method that interrupts computation at key convergence oscillation nodes, uses AI-predicted convergence fields, and restarts the solution is also a crucial component. This method skips the most time-consuming oscillation phase, thus accelerating the simulation. This invention not only provides the overall acceleration method but also includes the key prediction network structure, the interaction mechanism connecting the networks, and the specific applications of this method in fluid dynamics simulation, particularly in aero-engine design.
[0062] The core of this method is the use of a deep learning model based on the U-Net architecture as the prediction network, capable of predicting parameter convergence changes in numerical simulations. The connection network enables intelligent interaction with the numerical simulation software, providing real-time data extraction and writing, generating update vectors, and controlling the solver. The system does not blindly interrupt computation at arbitrary points in time; instead, it uses the connection network to monitor convergence indicators such as the solver's residual curves in real time, accurately identifying the "critical node" that marks the transition from a rapid decline phase to a prolonged oscillation phase in the convergence process. Identifying this node is crucial; stopping computation too early leaves the prediction network lacking sufficient convergence trend information, resulting in unreliable predictions; intervening too late wastes computation time and significantly reduces the acceleration effect. This invention locates this node by analyzing the high-frequency, small-amplitude oscillation characteristics of the residuals, ensuring the timeliness and effectiveness of the intervention.
[0063] This method does not replace the physical model and solution method of traditional numerical simulation software. Instead, it optimizes the solver's solution process by introducing an intelligent prediction framework that works in conjunction with existing software, thereby bypassing its unavoidable and time-consuming convergence oscillation phase. The overall architecture of this technical solution is as follows: Figure 1 As shown, it mainly consists of three core modules: a prediction network, a connection network, and a numerical simulation network. These three modules together form an intelligent closed loop of perception, prediction, and execution. Each part of the invention will be described in detail below with reference to the accompanying drawings.
[0064] Example 1
[0065] (1) The following is the overall technical route and workflow.
[0066] Before analyzing each module, let's first outline the overall workflow of this embodiment. The relationship between this embodiment and typical commercial software methods can be understood as the relationship between a winding mountain road and a tunnel. A typical solution process is similar to the original road; to solve a computational fluid dynamics problem, the process is often circuitous. This embodiment, however, establishes a tunnel between the computational starting point and the computational endpoint, thus avoiding the extensive twists and turns (or oscillations) required to reach the final computational goal.
[0067] like Figure 1 The specific construction and operation process is as follows:
[0068] S1: Configure the numerical simulation network and build the connection network
[0069] Configure the numerical simulation network, including parameters such as physical model, boundary conditions, and initial field; construct the network connection, including setting up data extraction channels, data writing channels, and constructing an update vector generation module, to realize intelligent interaction between the user and the numerical simulation network, perform real-time data extraction and writing, generate update vectors, and control the solver of the numerical simulation network.
[0070] S2: Construct the prediction network
[0071] A U-Net main network, comprising an encoder and a decoder, is constructed. The encoder downsamples the flow field grid parameters from multiple historical time steps to extract spatial features, while the decoder restores spatial resolution and reconstructs the flow field through deconvolution and skip connections. A spatiotemporal attention module is integrated into the U-Net. The spatial attention module is used to focus on key regions of the flow field, and the temporal attention module is used to learn the dynamic evolution law between the flow field states of multiple consecutive time steps. The network is trained using numerical simulation data. The flow field data from multiple time steps in the previous stage are used as input, and the corresponding near-converged flow field state is used as the target. This allows the network to learn the nonlinear mapping relationship from the initial stage of oscillation to the later stage of convergence, thereby outputting a predicted flow field with a complete physical structure, resulting in an optimized prediction network.
[0072] S3: Perform numerical simulations and calculations to accelerate convergence, such as... Figure 1 The specific steps are as follows:
[0073] S3.1: Initialization and Monitoring Phase:
[0074] S3.1.1 Users first set up the physical model, boundary conditions, and initial field in the numerical simulation network (such as CFX), and then start the calculation.
[0075] S3.1.2 At this point, the connection network is activated and begins monitoring the computational status of the numerical simulation network, especially determining whether the preset convergence oscillation key point has been reached. During this process, a data logger is used to record the grid node parameters obtained from the ten calculation steps preceding the current calculation step.
[0076] S3.2: Computation Interruption and Data Extraction Stage:
[0077] S3.2.1 When the solver of the numerical simulation network reaches a critical node that the program determines is in the early stage of convergence oscillation, specifically at the Nth step when the residual curve begins to enter high-frequency, small-amplitude oscillation, step S2.2 is executed; when the solver program determines that it is not a critical node in the early stage of convergence oscillation, it returns to step S1.2.
[0078] S3.2.2 The connection network sends a solver pause signal to the solver of the numerical simulation network. The solver maintains its current computation state, ensuring that physical quantities (such as pressure, velocity, temperature, etc.) on all mesh cells remain intact. Subsequently, the connection network's data extraction channel begins operation, extracting the mesh point parameters of all regions in the current iteration step and the ten steps prior to the current iteration step from the solver and summarizing them into a data format that can be input into the artificial intelligence prediction network.
[0079] S3.3: Accelerated Phase of Artificial Intelligence Prediction
[0080] S3.3.1: Artificial intelligence prediction network predicts parameters. The grid point parameters extracted in S2.2 are processed by the prediction network. This network is trained based on massive amounts of simulation data from real engineering applications. Its task is to predict the convergence value of the grid point parameters as they approach the convergence point. The specific prediction range depends on the complexity of the grid. The predicted parameters are the parameters of all points within the entire grid. These parameters are the grid point parameters predicted by the artificial intelligence network to approximate the stable solution in the traditional numerical simulation process. This process skips the most time-consuming step of hundreds of lengthy oscillations in the traditional numerical simulation calculation.
[0081] S3.3.2: Generation of Update Vector. This requires the use of a function within the connection network. The update vector generation module within the connection network calculates the predicted parameters and generates a new update vector (the update vector provides an iterative direction for the iterative calculation) to match the computational requirements of the numerical simulation network solver when restarting from a new state.
[0082] S3.3.3: Parameter Write-Back Numerical Simulation Solver. The grid point parameters predicted by artificial intelligence, along with the update vector, are written into the numerical simulation network solver via a connection network, overwriting the previous state. Finally, the connection network sends a computation restart signal to the solver, and the numerical simulation iterative calculation continues based on the obtained new parameters.
[0083] S3.4: Final Convergence and Output of Numerical Simulation Calculation
[0084] S3.4.1 In the computation monitor of the numerical simulation network, the user can determine whether the required computational accuracy has been achieved. If the required computational accuracy has been achieved, the computation can be terminated and the computational results can be output. If the accuracy is not achieved, the numerical simulation solver will continue to calculate until the requirements are met.
[0085] Specifically, the following is an explanation of the construction of the prediction network. Before the above operation process, the prediction network is trained based on the numerical simulation calculation results that have been tested and verified. The specific construction process is as follows, which includes two parts: building the artificial intelligence network architecture and training the artificial intelligence network.
[0086] The prediction network undertakes the crucial task of inferring a stable future convergence state from the oscillating current computational state. The construction process of the prediction network primarily considers the unique spatial complexity and temporal evolution characteristics of numerical simulation data. The construction process is as follows:
[0087] S2.1 Predictive Network Architecture Setup
[0088] The prediction network architecture used in this embodiment is based on the U-Net artificial intelligence structure. U-Net was originally designed for biomedical image segmentation, but it has been proven to perform well in tasks of capturing and reconstructing spatial features and data parameters in flow fields, effectively ensuring the physical structural integrity of the predicted flow field.
[0089] The encoder consists of multiple convolutional and pooling layers, responsible for downsampling the input current flow field mesh parameters (such as velocity and pressure fields). It identifies key structures, such as shock waves, separation regions, and main vortex structures, and then extracts abstract representations of these core features. Each downsampling step reduces spatial resolution but increases the number of feature channels, enhancing the accuracy of learning the flow field features.
[0090] The decoder gradually restores the spatial resolution through deconvolution operations. Each layer of the decoder receives feature maps from layers of the encoder at the same scale, achieving skip links. This ensures that when constructing the predicted flow field mesh, it can utilize both the high-level physical meaning extracted by the encoder and retain the fine spatial structure contained in the mesh structure from the finite element calculation, thereby generating prediction results with complete physical structure and data that are closer to the final convergence result.
[0091] However, the standard U-Net has limitations when handling temporal data. The convergence of numerical simulations is a dynamic process, with the current state closely related to historical states. Furthermore, this embodiment uses the results of ten iterative calculations before key nodes as input to the prediction network. Processing results from ten different time steps simultaneously is clearly unreasonable. To address this, this embodiment introduces a spatiotemporal attention mechanism. Spatial attention allows the network to more clearly identify the more important regions in the flow field. In computational fluid dynamics problems, researchers often have a preference for studying specific regions, which is often reflected in the grid density in engineering. The spatial attention module automatically assigns higher weights to these key regions, making the network prediction more focused. Temporal attention addresses the problem of inputting data from multiple consecutive time steps during training. The temporal attention mechanism can learn the evolutionary patterns between these continuous states, understanding the generation, development, and dissipation of vortices, or the forward and backward movement of shock waves. This allows the model's predictions to not only be based on a single convergence result but also to grasp the dynamic development of the entire convergence process.
[0092] S2.2, Artificial Intelligence Network Training
[0093] The training data for the prediction network consists of a large number of completed, high-precision numerical simulation examples. For each example, after the entire computation process is completed and convergence is achieved, the rapid descent region during convergence is first determined based on the number of grid cells, and the last ten steps are used as input features. The state of the same example nearing convergence is used as the training target value. In this way, the network learns the nonlinear mapping relationship from the initial stage of computational oscillation to the later stage of computational convergence. To improve the model's generalization ability, the training data is standardized, and various data augmentation techniques are employed, including perturbing the extracted parameters within a reasonable range to indirectly increase the sample size and enhance anti-interference ability, ensuring that the trained model has good adaptability to different operating conditions.
[0094] The following are instructions for building a connected network:
[0095] The unique connection network provided in this embodiment includes a data extraction channel and a data writing channel, as well as an update vector generation module. It also has the ability to control the solver to start and pause the solution process via a solver control language. The construction steps are as follows:
[0096] S1.1 Constructing data extraction and data writing channels
[0097] These two channels utilize the solver's native interface to extract input. The raw data format extracted from the solver cannot meet the input requirements of artificial intelligence networks, so it needs to be rearranged to meet the parameter classification requirements of different spaces, times, and categories, facilitating training and prediction.
[0098] S1.2 Update Vector Generation Module
[0099] The first thing to clarify is how the numerical simulation solver works.
[0100] In numerical simulation, the core of the solver is the numerical solution scheme. Commonly used numerical solution schemes include finite difference, finite element, spectral method and finite volume method. Generally speaking, the solution process of these methods is roughly the same, including the following steps.
[0101] a. Use a simple function to approximate the flow variable to be determined.
[0102] b. Substitute this approximate relationship into the continuous governing equations to form a discrete system of equations.
[0103] c. Solve the system of algebraic equations.
[0104] The main difference between various numerical solution schemes lies in the way the flow variables are approximated and the corresponding discretization process.
[0105] The solver used combines the discrete equations of all variables into a globally coupled system of equations. This system of equations can be expressed as: [equation follows].
[0106] (1)
[0107] In the above formula, [A] is a coefficient matrix. [Φ] is the correction vector, and [R] is the residual vector. After solving for the correction vector, the initial [Φ], which is the variable value at the beginning of this iteration, is updated. All solution variables are updated simultaneously in this step.
[0108] Therefore, after modifying the coefficient matrix, the update vector module needs to modify the corresponding correction vector to ensure consistency in the solution process. This update vector module employs a lightweight neural network. By training it to capture the relationship between the coefficient matrix and the correction vector, it achieves accurate estimation of the correction vector.
[0109] S1.3 Solver Control Interface
[0110] The control interface is responsible for sending "pause" and "continue" commands to the solver. Its implementation also relies on the existing solver's script control functionality. This interface ensures the accuracy and reliability of computation interruptions and restarts, guaranteeing a seamless transition from the new initial field when the solver resumes solving.
[0111] The following is a description of the numerical simulation network:
[0112] The numerical simulation network relies on existing numerical simulation programs and is identical to existing numerical simulation software in its main computational procedures. However, it implements new functions using the control language of existing numerical simulation software. Most importantly, it possesses the ability to output the grid point parameters for each step of the calculation and the ability to continue the calculation on modified grids. This numerical simulation network can output the physical parameters at all computational grid points in any specified iteration step. It can accept externally provided grid point parameters as new initial conditions and continue the calculation based on these conditions. This functionality can be achieved through the configuration files of existing commercial software, which will not be elaborated upon here.
[0113] In the field of numerical simulation in fluid mechanics, traditional numerical simulations, in order to improve and accelerate computational convergence, require a low-precision initial calculation to obtain a coarse initial flow field as input, followed by a high-precision calculation. In contrast, the accelerated convergence method provided in this embodiment only requires one calculation because the predicted flow field proposed by artificial intelligence more closely approximates the convergent flow field. The convergence time of this method is also significantly shorter than traditional methods. This accelerated convergence method saves computational operations and time, eliminates the need for initial field selection, and effectively avoids unnecessary errors caused by manual processing.
[0114] In some other embodiments, an accelerated convergence system for numerical simulation based on artificial intelligence prediction is provided, comprising:
[0115] The configuration of the numerical simulation network and the construction of the connection network module are used to obtain the configuration parameters of the numerical simulation network, including the physical model, boundary conditions and initial field; the construction of the connection network includes setting up data extraction channels, data writing channels and constructing an update vector generation module to realize intelligent interaction between the user and the numerical simulation network, perform real-time data extraction and writing, generate update vectors and control the solver of the numerical simulation network;
[0116] A prediction network module is constructed to build the main U-Net network, which includes an encoder and a decoder. The encoder downsamples the flow field grid parameters of multiple historical time steps to extract spatial features, and the decoder restores the spatial resolution to reconstruct the flow field through deconvolution and skip connections. A spatiotemporal attention module is integrated into the U-Net. The spatial attention module is used to focus on key regions of the flow field, and the temporal attention module is used to learn the dynamic evolution law between the flow field states of multiple consecutive time steps. The network is trained using numerical simulation data. The flow field data of multiple time steps in the previous stage are used as input, and the corresponding near-converged flow field state is used as the target. The network learns the nonlinear mapping relationship from the initial stage of oscillation to the later stage of convergence, thereby outputting a predicted flow field with a complete physical structure, resulting in an optimized prediction network.
[0117] A numerical simulation and accelerated convergence module is provided to enable the connected network to monitor the solution status of the numerical simulation network in real time, pause the solver at residual oscillation nodes, and extract multi-step flow field data. The multi-step flow field data is transmitted to the prediction network, which predicts the flow field parameters of the globally approximate convergent solution. An update vector is calculated based on the flow field parameters to obtain the prediction result. Finally, the connected network writes the prediction result back to the numerical simulation network, overwriting the current state of the solver and guiding it to restart the iteration based on a high-quality initial field, thereby achieving accelerated convergence.
[0118] In some other embodiments, a computer device is provided, characterized in that the computer device includes an input interface and an output interface, and further includes: a processor adapted to implement one or more instructions;
[0119] A computer storage medium storing one or more instructions, the one or more instructions being adapted to be loaded and executed by the processor for the numerical simulation accelerated convergence method described above.
[0120] In some other embodiments, a computer storage medium is provided that stores one or more instructions adapted for loading and execution by a processor of the numerical simulation accelerated convergence method described above.
[0121] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A numerical simulation acceleration convergence method based on artificial intelligence prediction, characterized in that: Includes the following steps: S1. Configure the numerical simulation network and build the connection network. The configuration parameters of the numerical simulation network are obtained, including the physical model, boundary conditions, and initial field. The construction of the network includes setting up data extraction channels, data writing channels, and building an update vector generation module to realize intelligent interaction between the user and the numerical simulation network, perform real-time data extraction and writing, generate update vectors, and control the solver of the numerical simulation network. S2, Constructing the prediction network A U-Net main network, comprising an encoder and a decoder, is constructed. The encoder downsamples the flow field grid parameters from multiple historical time steps to extract spatial features, while the decoder restores spatial resolution and reconstructs the flow field through deconvolution and skip connections. A spatiotemporal attention module is integrated into the U-Net. The spatial attention module focuses on key regions of the flow field, while the temporal attention module learns the dynamic evolution of flow field states across multiple consecutive time steps. The network is trained using numerical simulation data, with flow field data from multiple time steps in the previous stage as input and the corresponding near-converged flow field state as the target. This allows the network to learn the nonlinear mapping relationship from the initial stage of oscillation to the later stage of convergence, thereby outputting a predicted flow field with a complete physical structure, resulting in an optimized prediction network. S3, numerical simulation and computation accelerate convergence The connection network monitors the solution status of the numerical simulation network in real time, pauses the solver at the residual oscillation node, and extracts multi-step flow field data. The multi-step flow field data is transmitted to the prediction network, which predicts the flow field parameters of the globally approximate convergent solution. The update vector is calculated based on the flow field parameters to obtain the prediction result. Finally, the connection network writes the prediction result back to the numerical simulation network, overwriting the current state of the solver and guiding it to restart the iteration based on a high-quality initial field, thereby accelerating convergence.
2. The numerical simulation accelerated convergence method based on artificial intelligence prediction according to claim 1, characterized in that: Step S1 involves constructing the connection network, which includes the following steps: S1.1 Constructing data extraction and data writing channels Extraction and input functions are implemented using the solver's native interface. The solver is rearranged to meet the parameter classification requirements of different spaces, times, and categories, which facilitates training and prediction. S1.2, Constructing the Update Vector Generation Module Solvers used for numerical simulations employ various operating methods, including finite difference, finite element, spectral methods, and finite volume methods. These solvers combine the discrete equations of all variables into a globally coupled system of equations, expressed as follows: （1） In the above formula, [A] is a coefficient matrix. [Φ] is the correction vector, and [R] is the residual vector. After solving for the correction vector, the initial [Φ], which is the variable value at the beginning of this iteration, is updated. All solved variables are updated at the same time. A lightweight neural network is constructed, and by training it to capture the relationship between the coefficient matrix and the correction vector, the accurate estimation of the correction vector is achieved. S1.3 Solver Control Interface The control interface is implemented through the solver's script control function. The control interface sends "pause" and "continue" commands to the solver to ensure that the calculation is interrupted and restarted.
3. The numerical simulation accelerated convergence method based on artificial intelligence prediction according to claim 1, characterized in that: Step S2, constructing the prediction network, includes the following steps: S2.1 Network Architecture Setup The main network architecture is based on the U-Net artificial intelligence structure. The encoder consists of multiple convolutional and pooling layers, which are responsible for downsampling the current flow field mesh parameters, identifying the main structure, and then extracting abstract representations of these core features. Each downsampling step reduces the spatial resolution but increases the number of feature channels, enhancing the learning accuracy of the flow field features. The decoder gradually restores the spatial resolution through deconvolution operations. Each layer of the decoder receives feature maps from layers of the same scale as the encoder, achieving skip connections and generating prediction results with complete physical structure and data that are closer to the final convergence result. The prediction network is input by taking the results of ten iterative calculations before the key nodes. At the same time, a spatiotemporal attention mechanism is introduced. The spatial attention module automatically assigns higher weights to these key regions, making the network more focused during prediction. The temporal attention mechanism is used to learn the evolution law between continuous states and understand the generation, development and dissipation of vortices, or the dynamic processes such as the forward movement and hysteresis of shock waves. S2.2, Artificial Intelligence Network Training The training data for the prediction network consists of a large number of completed, high-precision numerical simulation examples. For each example, after the entire computation process is completed and convergence is achieved, the rapid descent region during convergence is first determined based on the number of grid cells, and the last ten steps are used as input features. The state of the same example nearing convergence is used as the training target value. In this way, the network learns the nonlinear mapping relationship from the initial stage of computational oscillation to the later stage of computational convergence. To improve the model's generalization ability, the training data is standardized, and various data augmentation techniques are employed, including perturbing the extracted parameters within a reasonable range to indirectly increase the sample size and enhance anti-interference ability, ensuring that the trained model has good adaptability to different operating conditions.
4. The numerical simulation accelerated convergence method based on artificial intelligence prediction according to claim 1, characterized in that: Step S3 includes the following steps: S3.1 Initialization and Monitoring Phase: S3.1.1 First, set the physical model, boundary conditions, and initial field in the numerical simulation network, and then start the calculation; S3.1.2 Connect the network to activate and start monitoring the computational status of the numerical simulation network, especially to determine whether the preset convergence oscillation key point has been reached; the data logger is used to record the grid node parameters obtained from the ten calculation steps prior to the current calculation step; S3.2, Calculation Interruption and Data Extraction Stage: S3.2.1 When the solver of the numerical simulation network reaches a critical node that the program determines is in the early stage of convergence oscillation, specifically at the Nth step when the residual curve begins to enter high-frequency, small-amplitude oscillation, step S3.2.2 is executed; when the solver program determines that it is not a critical node in the early stage of convergence oscillation, it returns to step S3.1.
2. S3.2.2 The connection network sends a solver calculation pause signal to the solver of the numerical simulation network; the solver maintains the current calculation state to ensure that the physical quantities on all grid cells remain intact; subsequently, the data extraction channel of the connection network starts working, extracting the grid point parameters of all regions in the current iteration step and the ten steps before the current iteration step from the solver, and summarizing them into the data format input to the artificial intelligence prediction network; S3.3, Artificial Intelligence Prediction Acceleration Stage S3.3.1 Artificial intelligence prediction network predicts parameters. The grid point parameters extracted in step S3.2.2 are processed by the prediction network. The prediction span of the prediction network depends on the complexity of the grid. The predicted parameters are the parameters of all points in the entire grid. S3.3.2 Calculation and generation of update vector: The update vector generation module inside the network calculates the predicted parameters and generates a new update vector. The update vector provides an iterative direction for the iterative calculation to match the computational requirements of the numerical simulation network solver when it restarts from a new state. The S3.3.3 parameter write-back numerical simulation solver writes the grid point parameters and update vectors predicted by artificial intelligence into the numerical simulation network solver through the connection network, overwriting the previous state; finally, the connection network sends a calculation restart signal to the solver, and the numerical simulation iterative calculation continues based on the obtained new parameters. S3.4 Numerical Simulation Calculation Final Convergence and Output In the computation monitor of the numerical simulation network, the user determines whether the required computational accuracy has been achieved. If the required accuracy is achieved, the computation is terminated and the results are output. If the required accuracy is not achieved, the numerical simulation solver continues to compute until the requirements are met.
5. A numerical simulation accelerated convergence system based on artificial intelligence prediction, characterized in that: include: The configuration of the numerical simulation network and the construction of the connection network module are used to obtain the configuration parameters of the numerical simulation network, including the physical model, boundary conditions and initial field; the construction of the connection network includes setting up data extraction channels, data writing channels and constructing an update vector generation module to realize intelligent interaction between the user and the numerical simulation network, perform real-time data extraction and writing, generate update vectors and control the solver of the numerical simulation network; A prediction network module is constructed to build the main U-Net network, which includes an encoder and a decoder. The encoder downsamples the flow field grid parameters of multiple historical time steps to extract spatial features, and the decoder restores the spatial resolution to reconstruct the flow field through deconvolution and skip connections. A spatiotemporal attention module is integrated into the U-Net. The spatial attention module is used to focus on key regions of the flow field, and the temporal attention module is used to learn the dynamic evolution law between the flow field states of multiple consecutive time steps. The network is trained using numerical simulation data. The flow field data of multiple time steps in the previous stage are used as input, and the corresponding near-converged flow field state is used as the target. The network learns the nonlinear mapping relationship from the initial stage of oscillation to the later stage of convergence, thereby outputting a predicted flow field with a complete physical structure, resulting in an optimized prediction network. The numerical simulation and computational acceleration convergence module is used to enable the connection network to monitor the solution status of the numerical simulation network in real time, pause the solver at the residual oscillation node, and extract multi-step flow field data; transmit the multi-step flow field data to the prediction network, predict the flow field parameters of the globally approximate convergent solution through the prediction network, calculate the update vector based on the flow field parameters, and obtain the prediction result; finally, the connection network writes the prediction result back to the numerical simulation network, overwrites the current state of the solver, and guides it to restart the iteration based on a high-quality initial field, thereby achieving accelerated convergence.
6. A computer device, characterized in that, The computer device includes an input interface and an output interface, and further includes: a processor adapted to implement one or more instructions; A computer storage medium storing one or more instructions, said one or more instructions being adapted to be loaded by the processor and executed as described in any one of claims 1 to 4, to accelerate the convergence of numerical simulations.
7. A computer storage medium storing one or more instructions, said one or more instructions being adapted to be loaded by a processor and executed as described in any one of claims 1 to 4.