Multi-physical flow prediction method based on multi-branch neural network

Abstract

The invention discloses a multi-physical flow prediction method based on a multi-branch neural network. The method comprises the following steps: constructing a lattice Boltzmann model; complex fluid flow is simulated through a lattice Boltzmann model, and multi-physical fluid flow data are collected; constructing different branch network models to respectively process different physical field information, and respectively extracting corresponding features for training learning; constructing a multi-branch network model in combination with the plurality of branch network models subjected to training learning; in the multi-branch network model, information flow paths are arranged among the plurality of branch network models and are used for carrying out feature coupling, and cross learning and sharing are carried out on features belonging to different branch network models; training the multi-branch network model through the acquired multi-physical fluid flow data; and performing multi-physical flow prediction through the trained multi-branch network model. According to the method, the problems of difficulty in dynamic modeling of the multi-physical system, long iteration time and low calculation efficiency of a traditional calculation method can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of flow prediction, and in particular to a multi-physical flow prediction method based on a multi-branch neural network. Background Art

[0002] Computational Fluid Dynamics (CFD) is an important method for simulating and analyzing complex fluids. Traditional CFD methods based on the Navier-Stokes equations have been widely used in microfluidic simulations. However, such methods also have some intractable problems. Firstly, it is difficult to incorporate microscopic effects, such as the dynamics of wetting and interfacial slip. Secondly, it is difficult to handle large deformations and complex boundary problems. In addition, the dynamic modeling and prediction of complex multi-physical field systems are also a key challenge, which is related to multiple engineering applications. The so-called multi-physics refers to considering and simulating the interactions of multiple physical phenomena or physical fields simultaneously in a simulation or calculation process. In practical engineering and scientific problems, more than one physical phenomenon is often involved, such as fluid mechanics, heat conduction, structural mechanics, electromagnetic fields, etc. These physical phenomena may be coupled and interact with each other, so multi-physical simulations are required to comprehensively consider these factors.

[0003] The Lattice Boltzmann Method (LBM) as a new computational fluid method has significant advantages in the field of multi-physical coupling. It calculates the flow field parameters by solving the Boltzmann equation of particle distribution. It describes the essence with particles, which brings convenience to the treatment of complex river basins, multiphase flows, and multi-physical flows. However, the computational efficiency of LBM has always been the biggest problem in its application. To address this problem, for example, Hennigh et al. proposed the Lat-Net model, which uses a neural network to compress the simulation time and memory usage to improve computational efficiency, and uses a convolutional autoencoder and residual connections to improve accuracy; some researchers have improved the computational efficiency by using a neural network to approximately replace the traditional neural network operator; in addition, Corbetta et al. learned the collision operator of LBM from data through deep learning, and approximated the update of the collision operator with a nine-direction distribution function. Although this method improves the accuracy of learning and prediction, the problem of computational efficiency still exists. Summary of the Invention

[0004] Aiming at the problems that the dynamic modeling of complex multi - physical systems and the action modes of external physical fields are complex and diverse, difficult to handle, and have low computational efficiency, the present invention provides a multi - physical flow prediction method based on a multi - branch neural network. By designing dedicated branch neural networks to separately learn the changes of different external physical fields and extract their features, the changes of the flow field can be learned and predicted by a neural network model combined with the LBM, which can better learn the change rules of the lattice Boltzmann method. Finally, the prediction result is obtained by comprehensively combining the results of both through a loss function. This method can overcome the problems of traditional calculation methods, such as difficult dynamic modeling of multi - physical systems, long iteration time, and low computational efficiency.

[0005] To achieve the above object, the technical solution provided by the present invention is as follows:

[0006] A multi - physical flow prediction method based on a multi - branch neural network, comprising:

[0007] Further, a lattice Boltzmann model is constructed;

[0008] The complex fluid flow is simulated by the lattice Boltzmann model, and multi - physical fluid flow data is collected;

[0009] Different branch network models are constructed to separately process different physical field information, and the corresponding features are respectively extracted for training and learning;

[0010] A multi - branch network model is constructed by combining multiple branch network models that have completed training and learning; in the multi - branch network model, there is an information flow path between multiple branch network models for feature coupling, so that the features belonging to different branch network models can be cross - learned and shared;

[0011] The multi - branch network model is trained with the collected multi - physical fluid flow data;

[0012] Multi - physical flow prediction is performed by the trained multi - branch network model.

[0013] Further, the control equation of the lattice Boltzmann model is as shown in the following formula:

[0014]

[0015] Among them, f i represents the distribution function of the lattice Boltzmann, represents the equilibrium distribution function, Δt represents a time step, m i = Mf i , while S, M, and M -1 are the relaxation matrix, transformation matrix, and its inverse matrix respectively, and R iΔt is a discrete source term representing the influence of the external force field, and the subscript i represents the discretization of the parameters.

[0016] Furthermore, the source term varies according to the different applied external force fields. The external force source term R exerted by the sound field on the flow field i is expressed as the following formula:

[0017]

[0018] where ξ is the discrete velocity space, u is the macroscopic velocity, F is the external force exerted by the applied external force field, ρ is the fluid density, and c s is the lattice sound speed, which has different values in different models; f i is the distribution function.

[0019] Furthermore, the external force source term R exerted by the temperature field on the flow field i (x, t) is expressed as the following formula:

[0020]

[0021] where φ is the temperature, t is the time, u is the fluid velocity vector, τ φ is the relaxation factor, ξ i is the discrete velocity ω i is the weight of the model.

[0022] Furthermore, in the process of simulating complex fluid flow through the lattice Boltzmann model, there are two core steps, namely collision and migration;

[0023] Collision:

[0024] Migration: f i (x + c i δ t , t + δ t ) = f' i (x, t)

[0025] f' i represents the distribution function after collision.

[0026] Furthermore, the branch network models included in the multi-branch network model are respectively the physics-informed neural network model and the bypass network model;

[0027] There is an information flow path between the physics-informed neural network model and the bypass network model for feature coupling, so that features belonging to different branch network models can be cross-learned and shared.

[0028] Further, the physical information neural network model includes four layers. The first layer is the input layer, which receives data. The second layer is the hidden layer, which processes data and extracts data features. The third layer is the output layer, which outputs prediction results. The fourth layer is the loss function. By inputting the prediction results into the loss function module, the loss function is continuously fitted to train its own model.

[0029] Further, the multi-branch network model obtains the final multi-physical flow prediction result by synthesizing the results of the physical information neural network model and the bypass network model through its own loss function.

[0030] Compared with the prior art, the principles and advantages of the present technical solution are as follows:

[0031] The present technical solution designs a dedicated branch neural network to separately learn the changes of different externally applied physical fields and extract their features. The changes of the flow field can be learned and predicted by a neural network model combined with LBM, which can better learn the change rules of the lattice Boltzmann method. Finally, the prediction result is obtained by synthesizing the results of the two through the loss function. The present technical solution can overcome the problems of traditional calculation methods, such as difficult dynamic modeling of multi-physical systems, long iteration time, and low calculation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the services required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 It is a principle flow chart of a multi-physical flow prediction method based on a multi-branch neural network of the present invention;

[0034] Figure 2 It is a structural diagram of a physical information neural network model;

[0035] Figure 3 It is a structural diagram of a multi-branch network model;

[0036] Figure 4 It is a structural diagram of a square cavity;

[0037] Figure 5 It is a flow field data diagram obtained after the physical information neural network model is processed;

[0038] Figure 6 It is a temperature field data diagram obtained after the bypass network model is processed;

[0039] Figure 7The figure shows the simulation results of the cavity fluid predicted by the multi-branch network model. Detailed implementation manners

[0040] The present invention will be further described below in conjunction with specific embodiments:

[0041] As Figure 1 shown, a multi-physical flow prediction method based on a multi-branch neural network according to this embodiment includes the following steps:

[0042] S1. Construct a lattice Boltzmann model;

[0043] The control equation of the lattice Boltzmann model is shown as follows:

[0044]

[0045] Among them, f i represents the distribution function of the lattice Boltzmann, represents the equilibrium distribution function, Δt represents a time step, m i = Mf i , and S, M and M -1 are respectively the relaxation matrix, the transformation matrix and its inverse matrix, R i Δt is the discrete source term, representing the influence of the external force field, and the subscript i represents the discretization of the parameter.

[0046] The source term varies differently according to the different external force fields applied;

[0047] Among them, the external force source term R i applied by the sound field to the flow field is expressed as follows:

[0048]

[0049] In the above formula, ξ is the discrete velocity space, u is the macroscopic velocity, F is the external force applied by the external force field, ρ is the fluid density, c s is the lattice sound speed, which has different values in different models; f i is the distribution function.

[0050] The external force source term R i (x, t) applied by the temperature field to the flow field is expressed as follows:

[0051]

[0052] In the above formula, φ is the temperature, t is the time, u is the fluid velocity vector, τ φ is the relaxation factor, ξ i is the discrete velocity ω i is the weight of the model.

[0053] In the process of simulating complex fluid flow through the lattice Boltzmann model, there are two core steps, namely collision and migration;

[0054] Collision:

[0055] Migration: f i (x + c i δ t , t + δ t ) = f' i (x, t)

[0056] f' i represents the distribution function after collision.

[0057] S2. Simulate complex fluid flow through the lattice Boltzmann model and collect multi - physical fluid flow data;

[0058] S3. Construct different branch network models to process different physical field information respectively, and extract corresponding features for training and learning;

[0059] Among them, the constructed branch network models are the physical information neural network model and the bypass network model respectively;

[0060] As Figure 2 shown, the physical information neural network model includes four layers. The first layer is the input layer, which accepts data; the second layer is the hidden layer, which processes data and extracts data features; the third layer is the output layer, which outputs the prediction result; the fourth layer is the loss function. By inputting the prediction result into the loss function module, the loss function is continuously fitted to train its own model.

[0061] S4. Combine the multiple branch network models completed in training and learning to construct a multi - branch network model, as Figure 3 shown; in the multi - branch network model, there is an information flow path between multiple branch network models for feature coupling, so that the features belonging to different branch network models can be cross - learned and shared;

[0062] S5. Train the multi - branch network model with the multi - physical fluid flow data collected;

[0063] S6. Perform multi - physical flow prediction through the trained multi - branch network model. Specifically, the multi - branch network model obtains the final multi - physical flow prediction result by synthesizing the results of the physical information neural network model and the bypass network model through its own loss function.

[0064] To prove the effectiveness of the method described in the present invention, the following experiment is carried out:

[0065] Specifically, a case of fluid flow induced by heat diffusion in a sealed square cavity:

[0066] The structure of the square cavity is as Figure 4 shown. The flow space is set as a 100x100 two-dimensional grid. The initial temperature field in the square cavity is set to T1 = 270 (unit: Kelvin), and the boundary of the square cavity is set to a constant temperature T2 = 350. This constitutes a boundary condition with a fixed temperature difference, which can drive the heat diffusion of the fluid.

[0067] The lattice Boltzmann method with a source term is used to simulate the heat diffusion flow problem in the square cavity. The core control equation is

[0068]

[0069] In this case, R i represents the force of the temperature field on the flow field, and the formula is as follows:

[0070]

[0071] In the above formula, φ is the temperature, t is the time, u is the fluid velocity vector, τ φ is the relaxation factor, ξ i is the discrete velocity ω i is the weight of the model.

[0072] Based on the setting of this control equation, initial conditions, and boundary conditions, the case data of the heat diffusion in the square cavity can be obtained, and the data set required for the branch neural network model can be made. Through the multi-branch network model, the situation of the stationary fluid flow affected by temperature can be obtained. Specifically:

[0073] Among them, by preprocessing the data set, it can be input into the multi-branch network model for training. The physical information neural network model processes and extracts the data features of the flow field, and the bypass network model processes the data features of the temperature field. The flow field data diagram obtained after the physical information neural network model processes is as Figure 5 shown. The temperature field data diagram obtained after the bypass network model processes is as Figure 6 shown.

[0074] Finally, through the data predicted by the two branch networks, continuously fitting the loss function to achieve the final prediction effect, the results of the subsequent uncompleted numerical experiments can be predicted, that is, the situation of the stationary fluid flow affected by temperature can be obtained through the multi-branch network model, as Figure 7 shown.

[0075] The present invention designs a dedicated branch neural network to separately learn the changes of different externally applied physical fields and extract their features. The changes in the flow field can be learned and predicted by a neural network model combined with LBM, which can better learn the variation law of the lattice Boltzmann method. Finally, the prediction result is obtained by comprehensively combining the results of both through a loss function. The present invention can overcome the problems of traditional calculation methods, such as difficult dynamic modeling of multi-physical systems, long iteration time, and low calculation efficiency.

[0076] The above-described embodiments are only the preferred embodiments of the present invention, and do not limit the scope of implementation of the present invention. Therefore, all changes made according to the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A multi-physics flow prediction method based on a multi-branch neural network, characterized in that: include: Construct a lattice Boltzmann model; Simulate complex fluid flows through the lattice Boltzmann model and collect multi-physics fluid flow data; Construct different branch network models to process different physical field information respectively, and extract corresponding features for training and learning respectively; A multi-branch network model is constructed by combining multiple branch network models that have been trained and learned. In the multi-branch network model, information flow paths are set between multiple branch network models for feature coupling, so that features belonging to different branch network models can be cross-learned and shared. The multi-branch network model is trained by using the collected multi-physics fluid flow data; Multi-physics flow prediction is performed through the trained multi-branch network model.

2. The multi-physics flow prediction method based on a multi-branch neural network according to claim 1 is characterized in that: The governing equation of the lattice Boltzmann model is as follows: Among them, f i represents the lattice Boltzmann distribution function, represents the equilibrium distribution function, Δt represents a time step, m i =Mf i , And S, M and M -1 Then they are the relaxation matrix, the transformation matrix and its inverse matrix, R i Δt is the discrete source term, which indicates the influence of the external force field, and the subscript i indicates the discretization of the parameters.

3. The multi-physics flow prediction method based on a multi-branch neural network according to claim 2 is characterized in that: The source term varies according to the external force field applied. The external force source term R applied by the acoustic field to the flow field is i It is expressed as the following formula: Among them, ξ is the discrete velocity space, u is the macroscopic velocity, F is the external force applied by the external force field, ρ is the fluid density, c s is the lattice sound speed, which has different values ​​in different models; f i is the distribution function.

4. The multi-physics flow prediction method based on a multi-branch neural network according to claim 2 is characterized in that: The external force source term R applied by the temperature field to the flow field i (x, t) is expressed as follows: Where φ is temperature, t is time, u is the fluid velocity vector, τ φ is the relaxation factor, ξ i is the discrete speed ω i is the weight of the model.

5. A multi-physics flow prediction method based on a multi-branch neural network according to any one of claims 2 to 4, characterized in that: The process of simulating complex fluid flows through the lattice Boltzmann model includes two core steps, namely collision and migration; collision: migration: f i (x+c i δ t ,t+δ t )=f' i (x,t) f' i represents the distribution function after the collision.

6. The multi-physics flow prediction method based on a multi-branch neural network according to claim 1 is characterized in that: The branch network models included in the multi-branch network model are respectively a physical information neural network model and a bypass network model; An information flow path is provided between the physical information neural network model and the bypass network model for feature coupling, so that features belonging to different branch network models can be cross-learned and shared.

7. The multi-physics flow prediction method based on a multi-branch neural network according to claim 6 is characterized in that: The physical information neural network model includes four layers, the first layer is the input layer, which receives data; The second layer is the hidden layer, which processes data and extracts data features; the third layer is the output layer, which outputs the prediction results; the fourth layer is the loss function, which continuously fits the loss function by inputting the prediction results into the loss function module to train its own model.

8. A multi-physics flow prediction method based on a multi-branch neural network according to claim 6 or 7, characterized in that: The multi-branch network model integrates the results of the physical information neural network model and the bypass network model through its own loss function to obtain the final multi-physics flow prediction results.

Images