Three-dimensional unsteady flow field prediction method and system based on physical information driven deep learning
By resampling the flow field physical quantities on the unstructured grid to a Cartesian grid and combining it with a 3D ResUnet network and a composite loss function, the problems of high computational resource consumption and insufficient accuracy in 3D unsteady flow field prediction are solved, and efficient and accurate flow field prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGAN UNIV
- Filing Date
- 2026-04-30
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies for predicting three-dimensional unsteady flow fields suffer from high computational resource consumption, high time costs, and insufficient prediction accuracy. They are particularly difficult to meet the requirements for high efficiency and high frequency of use in multivariate turbulence problems or scenarios that require frequent design iterations.
A deep learning method based on physical information is adopted to resample the physical quantities of the flow field on the unstructured grid to the Cartesian grid, and construct the historical window input through dimensionless processing. The method is then trained by combining a 3D ResUnet prediction network and a composite loss function to improve the accuracy and efficiency of flow field prediction.
It enables rapid and accurate prediction of three-dimensional unsteady flow fields, reduces the time cost of traditional CFD iterative solutions, preserves the dynamic evolution characteristics of the flow field, and improves prediction accuracy and engineering usability.
Smart Images
Figure CN122452431A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of flow field prediction technology, and relates to a method and system for predicting three-dimensional unsteady flow fields based on physical information-driven deep learning. Background Technology
[0002] Accurate prediction of three-dimensional unsteady flow fields is used for bridge wind engineering, flow analysis, and rapid assessment of complex external flow fields. While Computational Fluid Dynamics (CFD) simulations can provide high-precision pressure and velocity fields, they are time-consuming and computationally resource-intensive, making them unsuitable for efficient and high-frequency applications in scenarios involving multivariate turbulence or requiring frequent design iterations. Deep learning methods relying solely on pure data fitting for flow field prediction, although fast inference, suffer from insufficient prediction accuracy, large momentum residuals, and over-smoothing of local vortex structures, resulting in predictions lacking engineering usability. Summary of the Invention
[0003] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for predicting three-dimensional unsteady flow fields based on physical information-driven deep learning.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for predicting three-dimensional unsteady flow fields based on physics-information-driven deep learning, comprising the following steps: obtaining first intermediate data based on a three-dimensional CFD mesh file, an original result file, and external operating parameters; based on the first intermediate data, resampling the physical quantities at each time step and each mesh node on the unstructured mesh to a Cartesian mesh to obtain second intermediate data; based on the second intermediate data, dimensionlessizing the velocity and pressure fields, and constructing historical window inputs and supervision labels based on the dimensionlessized four-channel tensors of each time step; inputting the historical window inputs to a three-dimensional ResUnet prediction network to obtain the four-channel flow field prediction results for the next time step; constructing a composite loss function based on the four-channel flow field prediction results for the next time step and the supervision labels, updating the network parameters of the three-dimensional ResUnet prediction network based on the composite loss function, and outputting a three-dimensional unsteady flow field sequence based on the updated network parameters.
[0005] Furthermore, the composite loss function is :
[0006] in, This is the point value mean square error loss. For wavelet loss weights, For multi-level three-dimensional wavelet loss, The total weight of the physical constraint loss. For divergence loss, This represents the momentum residual loss.
[0007] Furthermore, the point value mean square error loss for:
[0008] in, For batch size, Number of output channels , Let be the number of discrete points in the three directions of the Cartesian grid. For batch ,aisle Spatial location ( The predicted value at ) For batch ,aisle Spatial location ( The truth value at () The multi-level three-dimensional wavelet loss for:
[0009] in, This represents the total number of levels. For wavelet decomposition level index, For the first Total number of wavelet coefficients at level 1 For the first Level 3 Haar wavelet transform operator, To predict the flow field, For true value flow fields; The divergence loss for:
[0010] in, For batch size, The number of internal grid points. For batch sample index , For 3D mesh points in , , Discrete indexes in direction, For the set of internal grid points, For velocity vector field, , For the first Each sample at the grid point velocity field divergence at a given location; The momentum residual loss for:
[0011] in, For batch size, The number of internal grid points. For batch sample index , For the set of internal grid points, for The residuals of the directional momentum equation at the corresponding grid points; for The residuals of the directional momentum equation at the corresponding grid points; for The residuals of the directional momentum equation at the corresponding grid points.
[0012] Furthermore, the velocity field divergence is :
[0013] in, For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for Twice the spacing between two adjacent grid points in the direction, used to construct the central difference discretization scheme. For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for Twice the spacing between two adjacent grid points in the direction, used to construct the central difference discretization scheme. For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for The spacing between two adjacent grids in the direction is twice the distance between them, which is used to construct the central difference discretization scheme.
[0014] Furthermore, the 3D ResUnet prediction network includes an encoder and a decoder. The encoder extracts multi-scale spatial features through two levels of 3D convolution and two downsampling operations, and the decoder restores the resolution through two levels of 3D transposed convolution.
[0015] Furthermore, the input for the history window is... :
[0016] in, For splicing operations, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, For the real number field, , Let be the number of discrete points in the three directions of the Cartesian grid.
[0017] Furthermore, the supervision label is :
[0018] in, for The flow field characteristic tensor at time t, For the real number field, , Let be the number of discrete points in the three directions of the Cartesian grid.
[0019] Furthermore, the external operating parameters include the characteristic velocity of the incoming flow, the fluid density, the characteristic length, and the kinematic viscosity.
[0020] Furthermore, the three-dimensional CFD mesh file is used to provide unstructured mesh node coordinates, cell topology relationships, boundary surface group information, and fluid physical quantities at each mesh node at each time step; the fluid physical quantities include pressure and velocity components in three spatial directions.
[0021] This invention also provides a three-dimensional unsteady flow field prediction system based on physics-information-driven deep learning, comprising: a reading module for obtaining first intermediate data from a three-dimensional CFD mesh file, original result file, and external operating parameters; a resampling module for resampling physical quantities at each time step and at each mesh node on an unstructured mesh to a Cartesian mesh based on the first intermediate data to obtain second intermediate data; a dimensionless modeling module for dimensionless modeling of the velocity and pressure fields based on the second intermediate data, and constructing historical window inputs and supervision labels based on the dimensionless four-channel tensors of each time step; a transmission module for transmitting the historical window inputs to a three-dimensional ResUnet prediction network to obtain the four-channel flow field prediction results for the next time step; and a prediction module for constructing a composite loss function based on the four-channel flow field prediction results for the next time step and the supervision labels, updating the network parameters of the three-dimensional ResUnet prediction network based on the composite loss function, and outputting a three-dimensional unsteady flow field sequence based on the updated network parameters.
[0022] Compared with the prior art, the present invention has the following beneficial technical effects: This invention presents a 3D unsteady flow field prediction method based on physically-informed deep learning. It resamples the flow field physical quantities from unstructured meshes to Cartesian meshes, eliminating the constraints of mesh topology differences on network input. The velocity and pressure fields are dimensionless, and historical input windows are constructed, preserving the dynamic evolution characteristics of the time dimension and improving the temporal modeling capability for unsteady flows. A 3D ResUnet prediction network is used to capture the spatial global features and local details of the flow field, enabling rapid inference of high-resolution 3D flow fields. A composite loss function incorporates physical consistency constraints into the network training process, improving prediction accuracy and avoiding non-physical oscillations that may occur with purely data-driven methods. This reduces the time cost of traditional CFD iterative solutions, enabling rapid and accurate prediction of 3D unsteady flow field sequences. Attached Figure Description
[0023] Figure 1 This is a flowchart of a three-dimensional unsteady flow field prediction method based on physical information-driven deep learning according to the present invention. Figure 2 This is a schematic diagram of a three-dimensional unsteady flow field prediction method based on physical information-driven deep learning according to the present invention. Detailed Implementation
[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0025] Example 1 This invention discloses a three-dimensional unsteady flow field prediction method based on physically-informed deep learning, such as... Figure 1 As shown, the process includes the following steps: reading the original 3D CFD result file and external operating parameters to obtain first intermediate data; based on the first intermediate data, resampling the physical quantities at each time step and each grid node on the unstructured grid to a Cartesian grid to obtain second intermediate data; based on the second intermediate data, dimensionlessizing the velocity field and pressure field, and constructing a historical window input and supervision label according to the dimensionlessized four-channel tensor of each time step; sending the historical window input to the 3D ResUnet prediction network to obtain the four-channel flow field prediction result for the next time step; constructing a composite loss function based on the four-channel flow field prediction result for the next time step and the supervision label, updating the network parameters of the 3D ResUnet prediction network according to the composite loss function, and outputting a 3D unsteady flow field sequence based on the updated network parameters.
[0026] The system reads the 3D CFD mesh file, the raw result file, and external operating parameters. The 3D CFD mesh file includes the geometric partitioning of the computational domain and network topology information. The raw result file provides the coordinates of unstructured mesh nodes, element topology relationships, boundary surface group information, and fluid physical quantities at each mesh node at each time step. These fluid physical quantities include pressure. and velocity components in three spatial directions External operating parameters include the characteristic velocity of the incoming flow. Fluid density Characteristic length and kinematic viscosity Furthermore, rule-based configuration parameters also need to be provided. and preset interpolation operators , Let be the number of discrete points in the three directions of the Cartesian grid. It is an interpolation operator that maps physical quantities of unstructured mesh nodes to a regular Cartesian mesh. The 3D CFD mesh file, the original result file, and external operating parameters are aggregated into first intermediate data. This first intermediate data undergoes regularization resampling and tensor organization. Unstructured mesh node coordinates, element topology relationships, and boundary surface group information are used to determine the original spatial extent. Used to determine regularized dimensions, such as Figure 2 As shown, a Cartesian mesh is constructed, which is as follows:
[0027] in, 、 、 To assign values to coordinates on a Cartesian grid, 、 、 For spatial indexing on a Cartesian grid, Let be the number of discrete points in the three directions of the Cartesian grid.
[0028] Mapping fluid physical quantities to a Cartesian grid :
[0029]
[0030] in, For the first A four-channel regular grid tensor at each time step, with the channel order fixed. ; For the first Dimensionless pressure field on a time-step regular grid; For the first On the time step rule grid Dimensionless velocity field in the direction; For the first On the time step rule grid Dimensionless velocity field in the direction; For the first On the time step rule grid Dimensionless velocity field in the direction; For tensor In channel 0, spatial location The value at that location corresponds to the dimensionless pressure. For tensor In Channel 1, Spatial Location The value at that location corresponds to Dimensionless velocity in direction; For tensor In Channel 2, Spatial Location The value at that location corresponds to Dimensionless velocity in direction; For tensor In Channel 3, Spatial Location The value at that location corresponds to Direction is dimensionless velocity.
[0031] Based on the second intermediate data, the velocity field and pressure field are dimensionless, and historical window input and supervision labels are constructed according to the dimensionless four-channel tensors of each time step.
[0032] Dimensionlessization:
[0033]
[0034]
[0035]
[0036] in, for Direction has a dimensionless velocity component. The characteristic velocity of the incoming flow. for Direction has a dimensionless velocity component. for Direction has a dimensionless velocity component. For fluid pressure, For fluid density, for Dimensionless velocity in direction for Dimensionless velocity in direction for Dimensionless velocity in direction It is a dimensionless pressure.
[0037] After dimensionless transformation, the input is no longer independent at each time step, but is organized into a fixed-length historical window. In this embodiment, the historical window length m=4 and the prediction step length k=1. The historical window input and supervision labels are constructed based on the four-channel tensors of each time step after dimensionless transformation.
[0038] Input in history window :
[0039] in, For splicing operations, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, For the real number field, , Let be the number of discrete points in the three directions of the Cartesian grid.
[0040] Supervision label is :
[0041] in, for The flow field characteristic tensor at time t, For the real number field, , Let be the number of discrete points in the three directions of the Cartesian grid.
[0042] The historical window input is fed into the 3D ResUnet prediction network to obtain the four-channel flow field prediction results for the next time step. The 3D ResUnet prediction network includes an encoder and a decoder. The encoder extracts multi-scale spatial features through two levels of 3D convolution and two downsampling operations, and the decoder restores the resolution through two levels of 3D transposed convolution.
[0043] The input to the 3D ResUnet prediction network is a historical window. The shape is 16×D×H×W, and the output prediction tensor is... The shape is 4×D×H×W, and the prediction tensor is... The output channel sequence is fixed at [p,u,v,w]
[0044] in, Forward mapping of the 3D ResUnet prediction network, This is the set of trainable parameters for the 3D ResUnet prediction network. Input a tensor for the history window.
[0045] A composite loss function is constructed based on the four-channel flow field prediction results of the next time step and the supervision labels.
[0046] The composite loss function is :
[0047] in, This is the point value mean square error loss. For wavelet loss weights, For multi-level three-dimensional wavelet loss, The total weight of the physical constraint loss. For divergence loss, This represents the momentum residual loss.
[0048] Point value mean square error loss for:
[0049] in, For batch size, Number of output channels , Let be the number of discrete points in the three directions of the Cartesian grid. For batch ,aisle Spatial location ( The predicted value at ) For batch ,aisle Spatial location ( The truth value at ().
[0050] Multi-level three-dimensional wavelet loss for:
[0051] in, This represents the total number of levels. For wavelet decomposition level index, For the first Total number of wavelet coefficients at level 1 For the first Level 3 Haar wavelet transform operator, To predict the flow field, For the true value flow field.
[0052] The divergence loss for:
[0053] in, For batch size, The number of internal grid points. For batch sample index , For 3D mesh points in , , Discrete indexes in direction, For the set of internal grid points, For velocity vector field, , For the first Each sample at the grid point The velocity field divergence at that location.
[0054] The velocity field divergence is :
[0055] in, For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for Twice the spacing between two adjacent grid points in the direction, used to construct the central difference discretization scheme. For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for Twice the spacing between two adjacent grid points in the direction, used to construct the central difference discretization scheme. For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for The spacing between two adjacent grids in the direction is twice the distance between them, which is used to construct the central difference discretization scheme.
[0056] The momentum residual loss for:
[0057] in, For batch size, The number of internal grid points. For batch sample index , For the set of internal grid points, for The residuals of the directional momentum equation at the corresponding grid points; for The residuals of the directional momentum equation at the corresponding grid points; for The residuals of the directional momentum equation at the corresponding grid points.
[0058]
[0059]
[0060]
[0061] in, for The partial derivative of directional velocity with respect to time, For velocity vector field, For gradient operators, for Directional convection term, for directional pressure gradient, Let Reynolds number be 1. for Laplace operator for directional velocity; for The partial derivative of directional velocity with respect to time, for Directional convection term, for directional pressure gradient, for Laplace operator for directional velocity; for The partial derivative of directional velocity with respect to time, for Directional convection term, for directional pressure gradient, for Laplace operator for directional velocity; It is the Reynolds number.
[0062] Update the network parameters of the 3D ResUnet prediction network based on the composite loss function. Based on the updated network parameters Output a three-dimensional unsteady flow field sequence.
[0063]
[0064] =
[0065] in, For the current step (the first step) The network prediction output tensor (step) has the following shape: ; For the next step (the first) (Step) Predict the new history window input tensor required, with shape as ; For the first The flow field characteristic tensor at time step; For the first The flow field characteristic tensor at time step; For the first The flow field characteristic tensor at time step; This indicates a splicing operation along the channel dimension.
[0066] Based on the current step prediction results Calculate the new historical window required for the next forecast. Autoregressive rolling prediction is performed to output a three-dimensional unsteady flow field sequence.
[0067] Example 2 The present invention also provides a three-dimensional unsteady flow field prediction system based on physical information-driven deep learning, including a reading module, a resampling module, a mapping module, a transport module, and a prediction module.
[0068] The system comprises the following modules: a reading module (for acquiring first intermediate data based on the 3D CFD mesh file, the original result file, and external operating parameters); a resampling module (for resampling physical quantities at each time step and at each mesh node on the unstructured mesh to a Cartesian mesh based on the first intermediate data to obtain second intermediate data); a dimensionless transformation module (for dimensionless transformation of the velocity and pressure fields based on the second intermediate data, and constructing historical window inputs and supervision labels based on the dimensionless four-channel tensors of each time step); a transmission module (for transmitting the historical window inputs to the 3D ResUnet prediction network to obtain the four-channel flow field prediction results for the next time step); and a prediction module (for constructing a composite loss function based on the four-channel flow field prediction results for the next time step and the supervision labels, updating the network parameters of the 3D ResUnet prediction network based on the composite loss function, and outputting a 3D unsteady flow field sequence based on the updated network parameters).
[0069] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
Claims
1. A method for predicting three-dimensional unsteady flow fields based on physically-informed deep learning, characterized in that, Includes the following steps: First intermediate data is obtained based on the 3D CFD mesh file, the original result file, and external operating parameters. Based on the first intermediate data, the physical quantities at each time step and each grid node on the unstructured grid are resampled to the Cartesian grid to obtain the second intermediate data; Based on the second intermediate data, the velocity field and pressure field are dimensionless, and historical window input and supervision labels are constructed according to the four-channel tensors of each time step after dimensionless transformation. The historical window input is fed into the three-dimensional ResUnet prediction network to obtain the four-channel flow field prediction results for the next time step. Based on the four-channel flow field prediction results of the next time step and the supervision labels, a composite loss function is constructed. The network parameters of the three-dimensional ResUnet prediction network are updated according to the composite loss function, and the three-dimensional unsteady flow field sequence is output based on the updated network parameters.
2. The three-dimensional unsteady flow field prediction method based on physically-informed deep learning according to claim 1, characterized in that: The composite loss function is: : in, This is the point value mean square error loss. For wavelet loss weights, For multi-level three-dimensional wavelet loss, The total weight of the physical constraint loss. For divergence loss, This represents the momentum residual loss.
3. The method for predicting three-dimensional unsteady flow fields based on physically-informed deep learning according to claim 2, characterized in that: The point value mean square error loss for: in, For batch size, Number of output channels , Let be the number of discrete points in the three directions of the Cartesian grid. For batch ,aisle Spatial location ( The predicted value at ) For batch ,aisle Spatial location ( The truth value at () The multi-level three-dimensional wavelet loss for: in, This represents the total number of levels. For wavelet decomposition level index, For the first Total number of wavelet coefficients at level 1 For the first Level 3 Haar wavelet transform operator, To predict the flow field, For true value flow fields; The divergence loss for: in, For batch size, The number of internal grid points. For batch sample index , For 3D mesh points in , , Discrete indexes in direction, For the set of internal grid points, For velocity vector field, , For the first Each sample at the grid point velocity field divergence at a given location; The momentum residual loss for: in, For batch size, The number of internal grid points. For batch sample index , For the set of internal grid points, for The residuals of the directional momentum equation at the corresponding grid points; for The residuals of the directional momentum equation at the corresponding grid points; for The residuals of the directional momentum equation at the corresponding grid points.
4. The three-dimensional unsteady flow field prediction method based on physically-informed deep learning according to claim 3, characterized in that: The velocity field divergence is : in, For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for Twice the spacing between two adjacent grid points in the direction, used to construct the central difference discretization scheme. For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for Twice the spacing between two adjacent grid points in the direction, used to construct the central difference discretization scheme. For grid points Place Directional velocity components The value of , For grid points Place Directional velocity components The value of , for The spacing between two adjacent grids in the direction is twice the distance between them, which is used to construct the central difference discretization scheme.
5. The three-dimensional unsteady flow field prediction method based on physically-informed deep learning according to claim 2, characterized in that: The 3D ResUnet prediction network includes an encoder and a decoder. The encoder extracts multi-scale spatial features through two levels of 3D convolution and two downsampling operations, and the decoder restores the resolution through two levels of 3D transposed convolution.
6. The method for predicting three-dimensional unsteady flow fields based on physically-informed deep learning according to claim 2, characterized in that: The historical window input is : in, For splicing operations, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, for The flow field characteristic tensor at time t, For the real number field, , Let be the number of discrete points in the three directions of the Cartesian grid.
7. The method for predicting three-dimensional unsteady flow fields based on physically-informed deep learning according to claim 6, characterized in that: The supervision label is : in, for The flow field characteristic tensor at time t, For the real number field, , Let be the number of discrete points in the three directions of the Cartesian grid.
8. The method for predicting three-dimensional unsteady flow fields based on physically-informed deep learning according to claim 1, characterized in that: The external operating parameters include the characteristic velocity of the incoming flow, fluid density, characteristic length, and kinematic viscosity.
9. The three-dimensional unsteady flow field prediction method based on physically-informed deep learning according to claim 8, characterized in that: The three-dimensional CFD mesh file is used to provide the coordinates of unstructured mesh nodes, element topology, boundary surface group information, and fluid physical quantities at each mesh node at each time step; The fluid physical quantities include pressure and velocity components in three spatial directions.
10. A three-dimensional unsteady flow field prediction system based on physically-informed deep learning, characterized in that: Reading module: Used to obtain the first intermediate data based on the 3D CFD mesh file, the original result file, and external working parameters; Resampling module: used to resample the physical quantities at each time step and each grid node on the unstructured grid to a Cartesian grid based on the first intermediate data to obtain the second intermediate data; Dimensioning module: used to perform dimensionless transformation on the velocity field and pressure field based on the second intermediate data, and construct historical window input and supervision labels based on the four-channel tensors of each time step after dimensionless transformation; Delivery module: used to send the historical window input to the three-dimensional ResUnet prediction network to obtain the four-channel flow field prediction results for the next time step; Prediction module: It is used to construct a composite loss function based on the four-channel flow field prediction results and supervision labels of the next time step, update the network parameters of the three-dimensional ResUnet prediction network according to the composite loss function, and output the three-dimensional unsteady flow field sequence based on the updated network parameters.