CN-121766159-B - Complex geometric flow field prediction method and system

CN121766159BCN 121766159 BCN121766159 BCN 121766159BCN-121766159-B

Abstract

The invention belongs to the technical field of computational fluid mechanics, and provides a complex geometric flow field prediction method and a complex geometric flow field prediction system, which are used for acquiring a fluid field of a complex geometric structure, sampling the inside and the boundary of the fluid field and constructing strong form physical constraint and boundary condition constraint; the method comprises the steps of generating spherical control volumes with different scales in a fluid domain, converting the volume integral in the control volume into flux integral on a boundary based on a steady-state incompressible Navier-Stokes equation by using a divergence theorem, constructing a weak form loss function of mass conservation and momentum conservation, approximating a flow field variable by using a neural network model, constructing a total loss function by combining strong form physical constraint, boundary condition constraint and weak form loss function, optimizing a network by using a two-stage training strategy, and generating a final flow field prediction result. The invention realizes the prediction of the high-precision and physical conservation of the convection field in the complex geometric structure.

Inventors

LV LIN
ZHANG WEIZHENG
Xie Xunjie
PAN HAO
DUAN XIAOWEI
Sun Bingteng
DU QIANG

Assignees

山东大学

Dates

Publication Date: 20260508
Application Date: 20260305

Claims (8)

1. A method for predicting a flow field of a flow channel of a complex geometry heat exchanger is characterized by comprising the following steps: Acquiring a fluid domain of a complex geometric heat exchanger runner structure, sampling the inside and the boundary of the fluid domain, and constructing strong form physical constraint and boundary condition constraint; generating spherical control volumes of different scales in a fluid domain, wherein the spherical control volumes comprise large-scale subdomains for long-range coupling, medium-scale subdomains distributed along a runner skeleton and small-scale subdomains for local correction; Based on a steady-state incompressible Navier-Stokes equation, converting the volume integral in the control volume into flux integral on the boundary by using a divergence theorem, and constructing a weak form loss function of mass conservation and momentum conservation; Approximating a flow field variable by using a neural network model, constructing a total loss function by combining strong form physical constraint, boundary condition constraint and the weak form loss function, and optimizing a network by adopting a two-stage training strategy to generate a final flow field prediction result; Sampling and retaining a central point in a geometric bounding box, generating a control volume covering a large area, and obtaining a large-scale subdomain for capturing long-range physical association; extracting a topological skeleton of the fluid domain, distributing control volumes along skeleton paths to obtain a mesoscale subdomain, so that the mesoscale subdomain covers a main fluid transportation channel, and ensuring effective propagation of physical information along a complex flow channel; Uniformly sampling in a fluid domain to generate a small-radius control volume, and obtaining a small-scale subdomain for capturing local fine geometric features and correcting local errors; the process of constructing a weak form loss function of conservation of mass and conservation of momentum includes, for any control volume Volume fraction constraint using the divergence theorem Conversion to surface integral constraints Wherein: The mass conservation weak form residual is obtained by calculating the integral of the velocity flux across the control volume boundary: Momentum conservation weak form residuals are integrated by calculating the resultant of momentum flux, pressure, and viscous forces across the control volume boundary: ; Wherein, the In order to be able to achieve a speed, In the case of a pressure force, the pressure, Is the normal vector of the surface of the object, Is the Reynolds number.
2. A method for predicting flow field of complex geometry heat exchanger according to claim 1, wherein obtaining fluid domain of complex geometry, sampling inside and boundary of fluid domain, and constructing strong physical constraint and boundary condition constraint comprises defining fluid domain Ω and boundary thereof for given complex geometry Omega, randomly sampling configuration points within a fluid domain Point-by-point residuals for computing standard steady-state incompressible Navier-Stokes equations while sampling on boundaries To apply boundary conditions.
3. The method for predicting flow field of complex geometry heat exchanger flow channel as claimed in claim 1, wherein in the process of generating mesoscale subdomains, geometric skeleton is extracted by adopting average curvature algorithm, and the control volume center is ensured to be positioned on topological central line of flow channel.
4. The method for predicting flow field of complex geometry heat exchanger as claimed in claim 1, wherein in the process of converting volume integral in control volume into flux integral on boundary by using divergence theorem, calculation of surface flux integral adopts gridless monte carlo sampling method, dividing control volume boundary into spherical surface part and truncated geometric boundary part, and respectively carrying out sampling and weighted summation.
5. The method for predicting the flow field of the complex geometry heat exchanger flow channel according to claim 1, wherein the process of optimizing the network by adopting a two-stage training strategy comprises the steps of enabling a strong form loss, a boundary loss and a weak form mass conservation loss at the first stage as a pre-training stage, quickly inhibiting mass leakage in a large range and establishing global flow field continuity; The second stage is a refinement stage, on the basis of the first stage, weak form momentum conservation loss is activated, flow field details are corrected, and fine solution of physical consistency is realized.
6. The method of claim 1, wherein the neural network model employs a fourier feature embedding layer to map spatial coordinates to high dimensional features to mitigate spectral bias and capture multi-scale flow field features.
7. A complex geometry heat exchanger flow field prediction system employing the method of claim 1, comprising: The fluid calculation domain construction module is configured to acquire a fluid domain of the complex geometric heat exchanger runner structure, sample the inside and the boundary of the fluid domain and construct strong physical constraint and boundary condition constraint; The multi-scale integral subdomain construction module is configured to generate spherical control volumes with different scales in a fluid domain, and comprises a large-scale subdomain for long-range coupling, a middle-scale subdomain distributed along a runner skeleton and a small-scale subdomain for local correction; A weak form conservation loss building block configured to transform the volume integral within the control volume into flux integral on the boundary using the divergence theorem based on a steady-state incompressible Navier-Stokes equation, building a weak form loss function of conservation of mass and conservation of momentum; The two-stage combined optimization module is configured to approximate a flow field variable by utilizing a neural network model, construct a total loss function by combining strong form physical constraint, boundary condition constraint and the weak form loss function, and optimize the network by adopting a two-stage training strategy to generate a final flow field prediction result.
8. An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, which when executed by the processor, perform the steps in the method of any of claims 1-6.

Description

Complex geometric flow field prediction method and system Technical Field The invention belongs to the technical field of intersection of computational fluid mechanics and deep learning, and particularly relates to a complex geometric flow field prediction method and system. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. In the fields of aerospace, automotive industry, chemical energy, etc., components such as heat exchangers typically comprise complex geometries, such as tricycles minimum curved (Triply Periodic Minimal Surfaces, TPMS) structures. Conventional computational fluid dynamics (Computational Fluid Dynamics, CFD) methods are time consuming and computationally expensive to generate grids when dealing with such complex geometries. In recent years, physical information neural networks (Physics-informed Neural Network, PINN) have received attention as a gridless method. However, the standard PINN method relies primarily on point-wise strong form partial differential equation (PARTIAL DIFFERENTIAL Equations, PDE) residual minimization. In areas of complex geometry, tortuous flow paths (e.g., heat exchanger internal flow paths), such local point-by-point constraints are difficult to effectively propagate global physical information, resulting in unstable gradients, mass conservation violations, and convergence difficulties. Existing improved methods, such as region-based sampling or traditional weak form methods, often require integration or cumbersome region decomposition depending on the grid, and are difficult to handle complex industrial-scale geometries in truly grid-free frames. Disclosure of Invention In order to solve the problems, the invention provides a complex geometric flow field prediction method and a complex geometric flow field prediction system, which adopt the integral conservation law based on the divergence theorem as weak form constraint and combine with a multi-scale control volume sampling strategy to realize the prediction of the high-precision and physical conservation of the flow field in a complex geometric structure. According to some embodiments, the present invention employs the following technical solutions: A complex geometric flow field prediction method comprises the following steps: acquiring a fluid domain with a complex geometric structure, sampling the inside and the boundary of the fluid domain, and constructing strong form physical constraint and boundary condition constraint; generating spherical control volumes of different scales in a fluid domain, wherein the spherical control volumes comprise large-scale subdomains for long-range coupling, medium-scale subdomains distributed along a runner skeleton and small-scale subdomains for local correction; Based on a steady-state incompressible Navier-Stokes equation, converting the volume integral in the control volume into flux integral on the boundary by using a divergence theorem, and constructing a weak form loss function of mass conservation and momentum conservation; And approximating the flow field variable by using the neural network model, constructing a total loss function by combining strong form physical constraint, boundary condition constraint and the weak form loss function, and optimizing the network by adopting a two-stage training strategy to generate a final flow field prediction result. Alternatively, the fluid domain of the complex geometry is acquired, sampled within and at boundaries of the fluid domain, and the process for constructing the strong form of physical constraints and boundary condition constraints includes, for a given complex geometry, defining the fluid domain Ω and its boundariesOmega, randomly sampling configuration points within a fluid domainPoint-by-point residuals for computing standard steady-state incompressible Navier-Stokes equations while sampling on boundariesTo apply boundary conditions. In an alternative embodiment, the process of generating spherical control volumes of different scales in a fluid domain comprises sampling and retaining a central point in the fluid domain in a geometric bounding box, generating a control volume covering a large area, and obtaining a large-scale subdomain for capturing long-range physical associations; extracting a topological skeleton of the fluid domain, distributing control volumes along skeleton paths to obtain a mesoscale subdomain, so that the mesoscale subdomain covers a main fluid transportation channel, and ensuring effective propagation of physical information along a complex flow channel; And uniformly sampling in the fluid domain to generate a small-radius control volume, and obtaining a small-scale subdomain for capturing local fine geometric features and correcting local errors. As a further defined embodiment, during the generation of the mesoscale subdomain, an average curvature algorithm is used to extra