CN115879335A - Fluid multi-physical-field parameter prediction method based on graph-generated neural network - Google Patents
Fluid multi-physical-field parameter prediction method based on graph-generated neural network Download PDFInfo
- Publication number
- CN115879335A CN115879335A CN202210034356.1A CN202210034356A CN115879335A CN 115879335 A CN115879335 A CN 115879335A CN 202210034356 A CN202210034356 A CN 202210034356A CN 115879335 A CN115879335 A CN 115879335A
- Authority
- CN
- China
- Prior art keywords
- fluid
- physical
- field
- graph
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a fluid multi-physical field parameter prediction method based on a graph generation neural network, which adopts finite element simulation software COMSOL to establish an unsteady flow fluid model, construct fluid multi-physical field data sets with different basin temperature fields and speed field information, and generate the neural network through a training graph to obtain a prediction model of the fluid multi-physical fields (speed fields, temperature fields and the like), namely, fluid multi-physical field parameters at the future moment are predicted according to the fluid multi-physical field data. The invention introduces the neural network generated by the graph to predict the unsteady flow fluid multi-physical field parameters, obviously improves the accuracy and timeliness of the fluid multi-physical field parameter prediction, is beneficial to the rapid modeling and accurate prediction of the complex fluid dynamic process, and provides support for the research of fluid dynamics and multi-physical processes and the application thereof in the industrial field.
Description
Technical Field
The invention relates to a fluid multi-physical-field parameter prediction method, in particular to a fluid multi-physical-field parameter prediction method based on a graph generation neural network, and belongs to the field of fluid multi-physical-field parameter prediction.
Background
The fluid multi-physical-field parameter simulation research plays an increasingly important role in the field of modern science and engineering, can overcome the defects of theoretical research and experimental research, and improves the scientific and engineering capability. The method has important research significance in various industries such as aerospace industry, ship and hydrodynamics industry, automobile industry, nuclear industry and the like.
At present, computational Fluid Dynamics (CFD) numerical simulation based on a Navier-Stokes (N-S) equation is one of common methods for studying fluid structure interaction problems of broad scholars, and the method can better simulate the complex flow problem in the interaction process of a structure and a flow field. CFD obtains physical field parameters of a fluid by solving high precision numerical values of basic physical equations of the flow field. However, the process of solving the N-S equations is a computationally intensive task, typically consumes a large amount of computing resources, and is computationally expensive.
In recent years, deep learning techniques have been rapidly developed and effectively applied in many fields. Deep learning can quickly obtain a calculation result by learning a training data set fitting data mapping relation, and has higher accuracy. During the CFD calculation, flow rate information and pressure information are stored, computed and matched in the form of a digital matrix. Therefore, the neural network has feasibility in the aspect of processing the flow field problem, and meanwhile, the calculation efficiency of CFD can be effectively improved as the neural network does not need to solve a control equation.
The graph generation neural network is a connection model, and the dependence relationship in the graph is obtained through the information transmission mode between nodes in the network, so that the dynamic process of the fluid can be accurately simulated. Compared with traditional vector or matrix data, the graph data conceptually adds the concepts of edges and the whole, so that the graph data can be used for establishing and simulating more complex models through the edges, is closer to actual physical processes and has better generalization than a traditional neural network. Therefore, for the problem that complex fluid multi-physical field calculation is slow, the prediction research of the fluid multi-physical field parameters based on the graph generation neural network has important research significance and application value.
Disclosure of Invention
In view of the above-mentioned shortcomings of the existing fluid multi-physical field parameter prediction technology, the invention provides a fluid multi-physical field parameter prediction method based on a graph generation neural network, which adopts finite element simulation software COMSOL to establish an unsteady flow fluid model, constructs fluid multi-physical field data sets with different basin temperature fields and speed field information, and generates a neural network through a training graph to obtain a prediction model of the fluid multi-physical field (speed field, temperature field and the like), namely, fluid multi-physical field parameters at a future moment are predicted according to the fluid multi-physical field data. The invention introduces the neural network generated by the graph to predict the unsteady flow fluid multi-physical field parameters, obviously improves the accuracy and timeliness of the fluid multi-physical field parameter prediction, is beneficial to the rapid modeling and accurate prediction of the complex fluid dynamic process, and provides support for the research of fluid dynamics and multi-physical processes and the application thereof in the industrial field.
The technical scheme adopted by the invention for solving the technical problem is as follows: a fluid multi-physical field parameter prediction method based on a graph generation neural network is characterized in that a finite element simulation software COMSOL is adopted to establish an unsteady flow fluid model, fluid multi-physical field data sets with different basin temperature fields and speed field information are established, a neural network is generated through a training graph to obtain a prediction model of fluid multi-physical fields (speed fields, temperature fields and the like), and fluid multi-physical field parameters at a future moment are predicted according to the fluid multi-physical field data. The fluid multi-physical field data set establishing steps are as follows:
step S1: establishing a geometric model of unsteady flow fluid in COMSOL software, defining unit type, material property of the unit, geometric property of the unit, connectivity of the unit, basis function, boundary condition and load of the unit, and dividing the established model into grids;
step S2: assembly solving, namely assembling the units into a combined equation set of the whole discrete domain, and solving the combined equation set by an iterative method to obtain a fluid multi-physical-field finite element modeling result;
and step S3: the obtained solution is analyzed and evaluated according to relevant criteria, and information is extracted through post-processing to obtain a calculation result of the unsteady flow fluid.
Finite element analysis is solved by replacing a complex problem with a simpler one. The solution domain is considered to consist of a number of small interconnected subdomains called finite elements, each element being assumed to have a suitable (simpler) approximate solution, and the solution to the problem is derived by deriving an overall satisfaction condition (e.g. structural equilibrium condition) for solving the domain. This solution is not an exact solution, but an approximate solution, since the actual problem is replaced by a simpler problem.
And step S4: and constructing sample data sets under different physical field parameters according to the finite element model result of the fluid multi-physical field, selecting fluid models with different parameters such as flow velocity fields of the flow fields, temperature field information and the like, and establishing a fluid multi-flow field data set.
Further, training a graph generation neural network by using training data to obtain a fluid multi-physical field (speed field, temperature field and the like) prediction model, wherein the algorithm steps of the graph generation neural network are as follows:
step S1: gridding the established fluid multi-physical field data set, wherein the fluid in each grid is called a fluid micelle;
step S2: carrying out refined modeling on the micro-clusters through graph embedding, and processing and calculating the evolution of physical field parameters among the micro-clusters to generate a final hidden variable;
in the figure, the characteristics of each micelle are described by nodes. Different micelles have different physical field characteristics, and the parameterized characteristics are organized into a one-dimensional vector format through the network as characteristic vectors of respective nodes. And finally, establishing a dependency relationship for the nodes by establishing links of edges among different nodes, wherein the characteristics of the edges can be the distance among the nodes or the information transmission coefficient of a physical field among the nodes, and the design of the characteristics needs to be analyzed according to the physical characteristics of data so as to determine a characteristic organization form which is most favorable for network operation. Processing and calculating the evolution of physical field parameters among the micelles, forming a sequence of hidden variables through N-step information transmission, and generating final hidden variables; the graph generation neural network adopts an averaging method, and uses the neural network to perform aggregation operation to acquire the information of the neighbor nodes, and the calculation method is as follows:
in the formula (I), the compound is shown in the specification,
graph embedding, x, representing node v at level 0 v Input feature vector representing a node>
Graph embedding representing node v at level k, σ represents a non-linear activation function, and->
Representing graph embedding of the average node v in the k-1 layer,
graph embedding representing node v in the previous layer.
And step S3: the decoder extracts dynamic information from the final hidden variables and outputs parameter predictions of the fluid multi-physics field at future times.
After the technical scheme is adopted, the invention has the beneficial effects that:
the invention provides a fluid multi-physical field parameter prediction method based on a graph generation neural network. The graph generation neural network is a connection model, and the dependence relationship in the graph is obtained through the information transmission mode between nodes in the network, so that the dynamic process of the fluid can be accurately simulated. Graph data conceptually adds an edge, whole concept compared to traditional vector or matrix data, which enables the building and simulation of more complex models through edges, more closely to actual physical processes, and has better generalization than traditional neural networks. Therefore, for the problem that the complex fluid flow field is slow in calculation, the prediction research based on the graph generation neural network has important research significance and application value.
Drawings
FIG. 1 is a schematic flow diagram of the process of the present invention.
FIG. 2 is a diagram of an update process of the graph-generated neural network in the method of the present invention.
FIG. 3 is a network architecture diagram illustrating the generation of a neural network in the method of the present invention.
Detailed Description
The following further describes the implementation of the present invention with reference to the accompanying drawings.
The method aims at solving the problems that a large amount of computing resources are occupied during the existing computational fluid dynamics numerical simulation computation, the computation time is long, and the like, and utilizes the graph generation neural network to accurately simulate the fluid dynamic process, overcomes the influence of the cognitive limitation of a fluid dynamic process physical model, and thus rapidly predicts the fluid multi-physical field parameters.
The invention relates to a fluid multi-physical field parameter prediction method based on a graph generation neural network, which is characterized in that the graph generation neural network is applied to fluid multi-physical field parameter prediction, and a fluid dynamic process physical model is simulated through the graph generation neural network; the method flow is as shown in figure 1, a finite element simulation software COMSOL is adopted to establish an unsteady flow fluid model, a fluid multi-physical field data set with different basin temperature fields and speed field information is established, a neural network is generated through a training diagram to obtain a prediction model of the fluid multi-physical field (speed field, temperature field and the like), namely, the fluid multi-physical field parameters at the future moment are predicted according to the fluid multi-physical field data. The method introduces the graph generation neural network to predict the unsteady flow fluid multi-physical-field parameters, obviously improves the accuracy and timeliness of the fluid multi-physical-field parameter prediction, is beneficial to the rapid modeling and accurate prediction of the complex fluid dynamic process, and provides support for the research of fluid dynamics and multi-physical processes and the application thereof in the industrial field.
Suppose X t E x is the state of the fluid multi-physical field at time t.
Representing the state trajectories of the same watershed at K time steps in continuous time. The present invention uses a simulator s: x → x simulates the evolution of this physical dynamics, i.e. causally mapping the current state to a future state. The simulator calculates kinetic information reflecting how the current state changes and uses this information to update the current state to a future state. The simulator is constructed by means of a parameterized function approximator d θ : x → Y, where θ can be trained and optimized, and Y ∈ Y represents the kinetic information, whose semantics are determined by the update mechanism. In the present invention, the update mechanism is the parameter conduction of the fluid multi-physical field and the flow of the micro-cluster. The update mechanism can be viewed as being based on the current status->
And d θ To predict a next state>
Is measured as a function of (c). The update process is shown in fig. 2.
The network structure diagram of the graph generating neural network in the method is shown in figure 3, the graph generating neural network establishes a structure based on nodes, edges and a graph, establishes a dependency relationship among the nodes with the edges as ties, is more favorable for representing a complex interaction relationship, is more close to an actual physical process, and has better generalization compared with the traditional neural network. A general graph network takes as input a graph and returns a graph as input. The input graph data is subjected to a series of sampling, convolution, pooling and other operations of the graph generation neural network, and nodes, edges or embedded layers of the graph are output. The graph generation neural network adopts an averaging method, and uses the neural network to perform aggregation operation to acquire the information of the neighbor nodes, and the calculation method is as follows:
in the formula (I), the compound is shown in the specification,
graph embedding, x, representing node v at level 0 v Input feature vector representing a node>
Graph embedding representing node v at level k, σ represents a non-linear activation function, and->
Representing graph embedding of the average node v in the k-1 layer,
graph embedding representing node v in the previous layer.
In summary, the fluid multi-physical-field parameter prediction method based on the graph-generated neural network provided by the invention has the following characteristics:
1) Compared with the traditional fluid multi-physical field parameter calculation method based on computational fluid dynamics numerical simulation, the fluid multi-physical field parameter calculation method based on computational fluid dynamics numerical simulation does not need to artificially assume a complex physical model, greatly improves the calculation speed, and can accurately simulate the dynamic process of the fluid;
2) Compared with the traditional machine learning algorithm, the method is closer to the actual physical process, so that the timeliness and the accuracy are better;
the above description is only a preferred embodiment of the method of the present invention, and it should be noted that several modifications can be made in the actual implementation without departing from the spirit of the solution and apparatus of the present invention.
Claims (3)
1. A fluid multi-physical field parameter prediction method based on a graph generation neural network is characterized in that a finite element simulation software COMSOL is adopted to establish an unsteady flow fluid model, fluid multi-physical field data sets with different basin temperature fields and speed field information are established, a neural network is generated through a training graph to obtain a prediction model of fluid multi-physical fields (speed fields, temperature fields and the like), and fluid multi-physical field parameters at a future moment are predicted according to the fluid multi-physical field data.
2. The method of claim 1, wherein the fluid multi-physics field data set is created by the steps of:
step S1: establishing a geometric model of unsteady flow fluid in COMSOL software, defining unit type, material property of the unit, geometric property of the unit, connectivity of the unit, basis function, boundary condition and load of the unit, and dividing the established model into grids;
step S2: assembly solving, namely assembling the units into a combined equation set of the whole discrete domain, and solving the combined equation set by an iterative method to obtain a fluid multi-physical-field finite element modeling result;
and step S3: the obtained solution is analyzed and evaluated according to relevant criteria, and information is extracted through post-processing to obtain a calculation result of the unsteady flow fluid.
Finite element analysis is solved by replacing a complex problem with a simpler one. The solution domain is considered to consist of a number of small interconnected subdomains called finite elements, each element being assumed to have a suitable (simpler) approximate solution, and the solution to the problem is derived by deriving an overall satisfaction condition (e.g. structural equilibrium condition) for solving the domain. This solution is not an exact solution, but an approximate solution, since the actual problem is replaced by a simpler problem.
And step S4: and constructing sample data sets under different physical field parameters according to the finite element model result of the fluid multi-physical field, selecting fluid models with different parameters of the basin flow velocity field, the temperature field information and the like, and establishing a fluid flow field data set.
3. The method for predicting the parameters of the fluid multi-physics field based on the graph generating neural network is characterized in that the graph generating neural network is trained by using training data to obtain a fluid multi-physics field (velocity field, temperature field and the like) prediction model:
step S1: gridding the established fluid multi-physical field data set, wherein the fluid in each grid is called a fluid micelle;
step S2: carrying out fine modeling on the micro-clusters through graph embedding, and processing and calculating evolution of multi-physical field parameters among the micro-clusters to generate final hidden variables;
in the figure, the characteristics of each micelle are described by nodes. Different micelles have different physical field characteristics, and the parameterized characteristics are organized into a one-dimensional vector format through the network as characteristic vectors of respective nodes. And finally, establishing a dependency relationship for the nodes by establishing links of edges among different nodes, wherein the characteristics of the edges can be the distance among the nodes or the information transmission coefficient of a physical field among the nodes, and the design of the characteristics needs to be analyzed according to the physical characteristics of data so as to determine a characteristic organization form which is most favorable for network operation. Processing and calculating the evolution of physical field parameters among the micelles, forming a sequence of hidden variables through N-step information transmission, and generating final hidden variables; the graph generation neural network adopts an averaging method, and uses the neural network to perform aggregation operation to acquire the information of the neighbor nodes, and the calculation method is as follows:
in the formula (I), the compound is shown in the specification,
graph embedding, x, representing node v at level 0 v An input feature vector representing a node>
Graph embedding representing node v at level k, σ represents a non-linear activation function, and->
Represents a graph embedding of an average node v in level k-1, and @>
Graph embedding representing node v in the previous layer.
And step S3: the decoder extracts dynamic information from the final hidden variables and outputs parameter predictions of the fluid multi-physics field at future times.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210034356.1A CN115879335A (en) | 2022-01-12 | 2022-01-12 | Fluid multi-physical-field parameter prediction method based on graph-generated neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210034356.1A CN115879335A (en) | 2022-01-12 | 2022-01-12 | Fluid multi-physical-field parameter prediction method based on graph-generated neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115879335A true CN115879335A (en) | 2023-03-31 |
Family
ID=85756878
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210034356.1A Pending CN115879335A (en) | 2022-01-12 | 2022-01-12 | Fluid multi-physical-field parameter prediction method based on graph-generated neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115879335A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116776135A (en) * | 2023-08-24 | 2023-09-19 | 之江实验室 | Physical field data prediction method and device based on neural network model |
| CN116777010A (en) * | 2023-08-25 | 2023-09-19 | 之江实验室 | Model training method and task execution method and device |
-
2022
- 2022-01-12 CN CN202210034356.1A patent/CN115879335A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116776135A (en) * | 2023-08-24 | 2023-09-19 | 之江实验室 | Physical field data prediction method and device based on neural network model |
| CN116776135B (en) * | 2023-08-24 | 2023-12-19 | 之江实验室 | Physical field data prediction method and device based on neural network model |
| CN116777010A (en) * | 2023-08-25 | 2023-09-19 | 之江实验室 | Model training method and task execution method and device |
| CN116777010B (en) * | 2023-08-25 | 2023-12-19 | 之江实验室 | Model training method and task execution method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110222828B (en) | Unsteady flow field prediction method based on hybrid deep neural network | |
| CN115879335A (en) | Fluid multi-physical-field parameter prediction method based on graph-generated neural network | |
| Gao et al. | Network community detection based on the physarum-inspired computational framework | |
| CN109726465B (en) | Three-dimensional non-adhesive low-speed streaming numerical simulation method based on non-structural curved edge grid | |
| JP7221062B2 (en) | Fluid analysis system, fluid analysis method, and fluid analysis program | |
| CN114818549B (en) | Method, system, equipment and medium for calculating fluid mechanics parameter of object | |
| Rumpfkeil et al. | A dynamic sampling method for kriging and cokriging surrogate models | |
| Luo et al. | Modified whale optimization algorithm for infinitive impulse response system identification | |
| CN115935834A (en) | History fitting method based on deep autoregressive network and continuous learning strategy | |
| CN113221475A (en) | Grid self-adaption method for high-precision flow field analysis | |
| White et al. | Fast neural network predictions from constrained aerodynamics datasets | |
| White et al. | Neural networks predict fluid dynamics solutions from tiny datasets | |
| CN115859781A (en) | Flow field prediction method based on attention and convolutional neural network codec | |
| Ghoreishi et al. | Gaussian process regression for bayesian fusion of multi-fidelity information sources | |
| Wong et al. | Graph neural network based surrogate model of physics simulations for geometry design | |
| CN114692523A (en) | Flow velocity prediction method of self-adaptive high-dimensional hydrodynamic equation based on graph convolution | |
| CN112800690A (en) | Underwater folding and unfolding mechanism parameter optimization method based on group intelligent optimization algorithm | |
| CN112949944A (en) | Underground water level intelligent prediction method and system based on space-time characteristics | |
| CN116230124A (en) | High-temperature alloy performance prediction method based on ANN and UMAT coupling micro-tissue damage | |
| Strönisch et al. | Flow field prediction on large variable sized 2D point clouds with graph convolution | |
| CN116628854A (en) | Wing section aerodynamic characteristic prediction method, system, electronic equipment and storage medium | |
| CN114692529A (en) | CFD high-dimensional response uncertainty quantification method and device, and computer equipment | |
| CN114510872A (en) | Cloud server aging prediction method based on self-attention mechanism DLSTM | |
| CN113420513A (en) | Underwater cylinder turbulent flow partition flow field prediction method based on deep learning | |
| Özalp et al. | Physics-informed long short-term memory for forecasting and reconstruction of chaos |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |