Large-scale and small-scale decoupling solution simulation method and system
Through the size-scale decoupling solution method based on spectral physical information neural network, combined with spectral decomposition and neural network, the problems of large calculation volume and low accuracy of traditional methods are solved, and efficient and high-precision multi-scale dynamics simulation is achieved, suitable for complex fluids Dynamics problem.
Patent Information
- Application Number
- CN202410870338.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-01
- Publication Date
- 2024-11-01
AI Technical Summary
Traditional numerical methods have high computational volume and low efficiency in multi-scale dynamic simulations. It is difficult for existing neural network methods to effectively deal with equation solutions for high-frequency or multi-scale structures, especially when spectral deviations exist.
Adopt the size-scale decoupling solution method based on spectral physical information neural network, and combine spectral decomposition and neural network to separate dynamic behaviors of different scales. Using the advantages of spectral methods and neural networks, we can realize efficient and high-precision simulation of multi-scale complex systems. . The specific steps include obtaining the initial data, mapping it to the spectral space, establishing a spectral neural network, training and fusing large and small scale information to obtain full-scale physical quantities.
It improves calculation efficiency and accuracy, can naturally adapt to different scales without relying on preset grids, greatly reduces calculation costs, reduces spectrum deviations, and significantly improves the performance of multi-scale simulations, and is suitable for complex fluid dynamics problems.
Smart Images

Figure CN118886296A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of multi-scale simulation of fluid mechanics, and in particular to a large-scale decoupling solution simulation method and system based on a spectral physics information neural network. Background Art
[0002] Multiscale phenomena are widely present in scientific fields such as physics, biology, engineering, and social systems. These phenomena involve patterns at multiple spatiotemporal scales and the close connections between them, making system analysis and prediction particularly difficult. The key to solving this complex problem lies in understanding the interaction between dynamic changes at large and small scales. Large-scale dynamics reflects the dominant patterns and main structures in the system from a macroscopic perspective, while small-scale dynamics focuses on exploring the microscopic components, details, and disturbances that can significantly affect the system as a whole. In-depth understanding and accurate prediction of dynamic changes at various spatiotemporal scales can not only improve our ability to model and simulate various phenomena, but also help us gain a deep insight into the emergent properties of the system, thus playing an important role in the design and optimization of new materials, biological processes, and technological systems.
[0003] Traditional numerical methods, including direct numerical simulation (DNS), large eddy simulation (LES) and lattice Boltzmann method (LBM), all have certain limitations when studying multiscale dynamics. The computational complexity of DNS increases significantly with the increase of system complexity, so it is not economical and practical in natural or engineering flow scenarios with high Reynolds numbers; LES still has difficulties in accurately simulating sub-grid scales; LBM requires a very large number of grids and a very small time step, which is time-consuming and increases computational costs. Existing technologies for enhancing small-scale simulation, such as multigrid method, regional decomposition method and adaptive mesh refinement, can concentrate computing resources in local areas, but they still face challenges in improving the resolution and computational efficiency of the entire range.
[0004] In recent years, with the widespread application of neural networks, more and more researchers have begun to try to use neural networks to explore multiscale dynamics, such as modeling the Reynolds stress tensor in the Reynolds averaged equations (RANS) and simulating sub-grid scales in LES. However, these methods still face the problems of sensitivity to noise amplification and huge computational complexity when processing high-resolution data. On the other hand, for partial differential equations related to multiscale dynamics, neural networks are regarded as a new type of solution tool. Methods such as physical information neural networks (PINN) and deep generative models (DGM) can integrate control equations and boundary conditions into loss functions or neural network architectures to obtain approximate solutions. However, for solutions to equations with high frequency or multi-scale structures caused by spectral deviations, there is still a lack of technology that can effectively deal with related problems. Summary of the invention
[0005] The purpose of the present invention is to provide a large-scale decoupling solution simulation method and system based on spectral physical information neural network, which can separate the dynamic behaviors of different scales and integrate the advantages of spectral method and neural network to achieve efficient and high-precision simulation of multi-scale complex systems.
[0006] The purpose of the present invention can be achieved by the following technical solutions:
[0007] A method for solving the large-scale decoupling simulation based on a spectral physics information neural network comprises the following steps:
[0008] Step 1: Obtain initial data of the physical quantity to be solved;
[0009] Step 2: Map the initial data to the spectral space, expand it with the basis functions of the spectral space, retain the main part of the spectral space information, and obtain the large-scale physical quantity;
[0010] Step 3: Take time as input and the spectral spatial information of small-scale physical quantities as output to establish a spectral neural network architecture;
[0011] Step 4: Based on the physical equations, the equations satisfied by the small-scale spectral information are derived to train the spectral neural network;
[0012] Step 5: Fusing the large-scale physical quantity with the small-scale physical quantity in the spectral space output by the trained spectral neural network to obtain the full-scale physical quantity in the spectral space;
[0013] Step 6: Transform the full-scale physical quantity in the spectral space back to the physical space to obtain the full-scale physical quantity.
[0014] The specific step 2 is: is the basis function of the spectral space, and the initial data u(t,x) is expressed as follows in the spectral space basis:
[0015]
[0016] Among them, the coefficient U k (t)=(u(t,x),φ k (x) represents the projection information of the initial data in the spectrum space, x, t represent the spatial coordinate and time coordinate respectively;
[0017] The main part of the spectral space information, namely the first M spectral information, is retained to obtain large-scale physical quantities.
[0018] In step 3, the input and output of the spectral neural network satisfy the following approximate relationship:
[0019]
[0020] in, It is the spectral space information of small-scale physical quantities, neural network Defines the mapping relationship between a single input time t and the small-scale spectral information output, represents a high-dimensional space consisting of positive real numbers and a real vector space of dimension |θ|, is the complementary space of the large-scale spectral space, is a collection of large-scale physical quantity information, represents a preset activation function, and θ represents the set of all parameters to be trained, including weights and offsets.
[0021] The loss function used by the spectral neural network is as follows:
[0022]
[0023] in, is the loss function of the initial value, which is obtained by projecting the initial value into the spectral space and separating the initial value of the small-scale spectral information; is the partial differential equation loss function satisfied by the small-scale information derived from the physical equation; w PDE and w IC The weights of the partial differential equation loss function and the initial value loss function are and denote spectral space transform and inverse transform respectively; Represents nonlinear partial differential operators in partial differential equations; L represents the error 2 norm, and They represent the true initial conditions and the initial conditions of the spectral neural network output under the parameter θ respectively.
[0024] The small-scale physical quantity and the large-scale physical quantity in the spectral space output by the spectral neural network after the fusion training are used to obtain the full-scale physical quantity in the spectral space:
[0025]
[0026] in, is a large-scale physical quantity, is the small-scale physical quantity output by the spectral neural network under the parameter θ.
[0027] A large-scale decoupling solution simulation system based on spectral physical information neural network, comprising:
[0028] Initial data acquisition module: obtains the initial data of the physical quantity to be solved;
[0029] Large-scale physical quantity decomposition module: maps the initial data to the spectral space, expands it with the basis functions of the spectral space, retains the main part of the spectral space information, and obtains the large-scale physical quantity;
[0030] Spectral neural network construction and training module: Take time as input and spectral space information of small-scale physical quantities as output to establish the spectral neural network architecture; deduce the equations satisfied by small-scale spectral information based on physical equations and train the spectral neural network;
[0031] Fusion module: Fusion of large-scale physical quantities with small-scale physical quantities in the spectral space output by the trained spectral neural network to obtain full-scale physical quantities in the spectral space;
[0032] Output module: transform the full-scale physical quantity in the spectral space back to the physical space to obtain the full-scale physical quantity.
[0033] The size-scale physical quantity decomposition module performs the following operations: is the basis function of the spectral space, and the initial data u(t,x) is expressed as follows in the spectral space basis:
[0034]
[0035] Among them, the coefficient U k (t)=(u(t,x),φ k (x) represents the projection information of the initial data in the spectrum space, x, t represent the spatial coordinate and time coordinate respectively;
[0036] The main part of the spectral space information, namely the first M spectral information, is retained to obtain large-scale physical quantities.
[0037] The input and output of the spectral neural network satisfy the following approximate relationship:
[0038]
[0039] in, It is the spectral space information of small-scale physical quantities, neural network Defines the mapping relationship between a single input time t and the small-scale spectral information output, represents a high-dimensional space consisting of positive real numbers and a real vector space of dimension |θ|, is the complementary space of the large-scale spectral space, is a collection of large-scale physical quantity information, represents a preset activation function, and θ represents the set of all parameters to be trained, including weights and offsets.
[0040] The loss function used by the spectral neural network is as follows:
[0041]
[0042] in, is the loss function of the initial value, which is obtained by projecting the initial value into the spectral space and separating the initial value of the small-scale spectral information; is the partial differential equation loss function satisfied by the small-scale information derived from the physical equation; w PDE and w IC are the weights of the partial differential equation loss function and the initial value loss function respectively; and denote spectral space transform and inverse transform respectively; Represents nonlinear partial differential operators in partial differential equations; L represents the error 2 norm, and They represent the true initial conditions and the initial conditions of the spectral neural network output under the parameter θ respectively.
[0043] The small-scale physical quantity and the large-scale physical quantity in the spectral space output by the spectral neural network after the fusion training are used to obtain the full-scale physical quantity in the spectral space:
[0044]
[0045] in, is a large-scale physical quantity, is the small-scale physical quantity output by the spectral neural network under the parameter θ.
[0046] Compared with the prior art, the present invention has the following beneficial effects:
[0047] First, the present invention proposes for the first time a method of combining physical information neural network and spectral decomposition to construct a spectral physical information neural network (SpectralPINN) framework to achieve modeling of small-scale dynamic systems. In addition, the scheme decomposes large and small scales, independently simulates large scales, and treats small scales as subordinate systems, so that complex problems in multi-scale dynamics can be handled more efficiently and conveniently in industrial computing.
[0048] Second, the previous small-scale simulation technology cannot always guarantee both computational efficiency and computational accuracy. The present invention proposes a solution for decoupling large and small scales, which can not only improve computational efficiency but also ensure computational accuracy. Specifically, firstly, large-scale data is obtained using a relatively low-computation technology, and then SpectralPINN is used to refine the small-scale structure. The small scale is solved by PINN, which improves computational efficiency; the partial differential equation is embedded in the loss function by PINN, which ensures the accuracy of the calculation.
[0049] Third, on the one hand, the present invention gives full play to the advantages of PINN in multi-scale simulation compared with traditional methods: it can naturally adapt to different scales within the system without relying on preset grids, greatly reducing the need for refined grids and related computational costs, and improving computational efficiency without sacrificing computational accuracy. On the other hand, the present invention also makes up for some of the shortcomings of PINN: in order to reduce the spectral deviation associated with PINN, the equations of small-scale dynamics are directly derived and learned in the spectral space, especially using Fourier space to represent functions of different scales, thereby improving the overall performance of the model.
[0050] Fourth, the present invention draws on the advantages of both deep learning and first principles, while avoiding the defects of systems that rely solely on data-driven, providing more possibilities for reliable and effective prediction of complex multi-scale systems. Numerical experiments have shown that this method is effective and versatile in solving complex fluid dynamics problems such as the one-dimensional Kuramoto-Sivashinsky equation and the two- and three-dimensional Navier-Stokes equation. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 is a flow chart of the method of the present invention;
[0052] Figure 2 It is a schematic diagram of the solution result of the one-dimensional Kuramoto-Sivashinsky equation of the present invention.
[0053] Figure 3 It is a schematic diagram of the solution results of the two-dimensional Navier-Stokes equation of the present invention.
[0054] Figure 4 It is a schematic diagram of the solution results of the three-dimensional Navier-Stokes equations of the present invention. DETAILED DESCRIPTION
[0055] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0056] This embodiment provides a large-scale decoupling solution simulation method based on spectral physical information neural network, such as Figure 1 As shown, the following steps are included:
[0057] Step 1: Obtain the initial data of the physical quantity to be solved, such as velocity, pressure, etc., through experiments or simulation methods such as RANS.
[0058] Step 2: Map the initial data to the spectral space, expand it with the basis functions of the spectral space, retain the main part of the spectral space information, and obtain large-scale physical quantities.
[0059] remember is the basis function of the spectral space, and the initial data u(t,x) is expressed as follows in the spectral space basis:
[0060]
[0061] Among them, the coefficient U k (t)=(u(t,x),φ k (x) represents the projection information of the initial data in the spectrum space, x, t represent the spatial coordinate and time coordinate respectively;
[0062] The main part of the spectral space information, namely the first M spectral information, is retained to obtain large-scale physical quantities.
[0063] Step 3: Take time as input and the spectral space information of small-scale physical quantities as output to establish a spectral neural network architecture.
[0064] The input and output of the spectral neural network satisfy the following approximate relationship:
[0065]
[0066] in, It is the spectral space information of small-scale physical quantities, neural network Defines the mapping relationship between a single input time t and the small-scale spectral information output, represents a high-dimensional space consisting of positive real numbers and a real vector space of dimension |θ|, is the complementary space of the large-scale spectral space, is a collection of large-scale physical quantity information, represents a preset activation function, and θ represents the set of all parameters to be trained, including weights and offsets.
[0067] Step 4: Based on the physical equations, derive the equations satisfied by the small-scale spectral information and train the spectral neural network.
[0068] The loss function used by the spectral neural network is as follows:
[0069]
[0070] in, is the loss function of the initial value, which is obtained by projecting the initial value into the spectral space and separating the initial value of the small-scale spectral information; is the partial differential equation loss function satisfied by the small-scale information derived from the physical equation; w PDEand w IC are the weights of the partial differential equation loss function and the initial value loss function respectively; and denote spectral space transform and inverse transform respectively; Represents nonlinear partial differential operators in partial differential equations; L represents the error 2 norm, and They represent the true initial conditions and the initial conditions of the spectral neural network output under the parameter θ respectively.
[0071] Step 5: Fuse the large-scale physical quantity with the small-scale physical quantity in the spectral space output by the trained spectral neural network to obtain the full-scale physical quantity in the spectral space:
[0072]
[0073] in, is a large-scale physical quantity, is the small-scale physical quantity output by the spectral neural network under the parameter θ.
[0074] Step 6: Transform the full-scale physical quantity in the spectral space back to the physical space to obtain the full-scale physical quantity:
[0075] u θ (t=U θ (t)φ
[0076] The above method is used to solve different partial differential equations respectively, and the results are as follows: Figure 2-4 As shown, the results of small-scale data obtained by this method in Fourier space and physical space are almost consistent with the numerical simulation results, demonstrating the effectiveness of this method in solving multi-scale problems.
[0077] This embodiment also provides a large-scale decoupling solution simulation system based on a spectral physics information neural network, including:
[0078] Initial data acquisition module: obtains the initial data of the physical quantity to be solved;
[0079] Large-scale physical quantity decomposition module: maps the initial data to the spectral space and uses the basis function of the spectral space to expand the representation, retaining the main part of the spectral space information to obtain the large-scale physical quantity;
[0080] Spectral neural network construction and training module: Take time as input and spectral space information of small-scale physical quantities as output to establish the spectral neural network architecture; deduce the equations satisfied by small-scale spectral information based on physical equations and train the spectral neural network;
[0081] Fusion module: Fusion of large-scale physical quantities with small-scale physical quantities in the spectral space output by the trained spectral neural network to obtain full-scale physical quantities in the spectral space;
[0082] Output module: transform the full-scale physical quantity in the spectral space back to the physical space to obtain the full-scale physical quantity.
[0083] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.
[0084] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A method for solving the size-scale decoupling simulation based on spectral physics information neural network, characterized in that: The following steps are involved: Step 1: Obtain initial data of the physical quantity to be solved; Step 2: Map the initial data to the spectral space, expand it with the basis functions of the spectral space, retain the main part of the spectral space information, and obtain the large-scale physical quantity; Step 3: Take time as input and the spectral spatial information of small-scale physical quantities as output to establish a spectral neural network architecture; Step 4: Based on the physical equations, the equations satisfied by the small-scale spectral information are derived to train the spectral neural network; Step 5: Fusing the large-scale physical quantity with the small-scale physical quantity in the spectral space output by the trained spectral neural network to obtain the full-scale physical quantity in the spectral space; Step 6: Transform the full-scale physical quantity in the spectral space back to the physical space to obtain the full-scale physical quantity.
2. According to claim 1, a method for solving the size-scale decoupling simulation based on spectral physics information neural network is characterized in that: The specific step 2 is: is the basis function of the spectral space, and the initial data u(t,x) is expressed as follows in the spectral space basis: Among them, the coefficient U k (t)=(u(t,x),φ k (x)) represents the projection information of the initial data in the spectrum space, x, t represent the spatial coordinate and time coordinate respectively; The main part of the spectral space information, namely the first M spectral information, is retained to obtain large-scale physical quantities.
3. The method for solving the size-scale decoupling simulation based on spectral physics information neural network according to claim 1 is characterized in that: In step 3, the input and output of the spectral neural network satisfy the following approximate relationship: in, It is the spectral space information of small-scale physical quantities, neural network Defines the mapping relationship between a single input time t and the small-scale spectral information output, represents a high-dimensional space consisting of positive real numbers and a real vector space of dimension |θ|, is the complementary space of the large-scale spectral space, is a collection of large-scale physical quantity information, represents a preset activation function, and θ represents the set of all parameters to be trained, including weights and offsets.
4. The method for solving the size-scale decoupling simulation based on spectral physics information neural network according to claim 3 is characterized in that: The loss function used by the spectral neural network is as follows: in, is the loss function of the initial value, which is obtained by projecting the initial value into the spectral space and separating the initial value of the small-scale spectral information; is the partial differential equation loss function satisfied by the small-scale information derived from the physical equation; w PDE and w IC are the weights of the partial differential equation loss function and the initial value loss function respectively; and denote the spectral space transform and inverse transform respectively; Represents nonlinear partial differential operators in partial differential equations; represents the L2 norm of the error, and They represent the true initial conditions and the initial conditions of the spectral neural network output under the parameter θ respectively.
5. The method for solving the size-scale decoupling simulation based on spectral physics information neural network according to claim 1 is characterized in that: The small-scale physical quantity and the large-scale physical quantity in the spectral space output by the spectral neural network after the fusion training are used to obtain the full-scale physical quantity in the spectral space: in, is a large-scale physical quantity, is the small-scale physical quantity output by the spectral neural network under the parameter θ.
6. A large-scale decoupling solution simulation system based on spectral physical information neural network, characterized in that: include: Initial data acquisition module: obtains the initial data of the physical quantity to be solved; Large-scale physical quantity decomposition module: maps the initial data to the spectral space, expands it with the basis functions of the spectral space, retains the main part of the spectral space information, and obtains the large-scale physical quantity; Spectral neural network construction and training module: takes time as input and spectral spatial information of small-scale physical quantities as output to establish a spectral neural network architecture; Based on the physical equations, the equations satisfied by the small-scale spectral information are derived to train the spectral neural network; Fusion module: Fusion of large-scale physical quantities with small-scale physical quantities in the spectral space output by the trained spectral neural network to obtain full-scale physical quantities in the spectral space; Output module: transform the full-scale physical quantity in the spectral space back to the physical space to obtain the full-scale physical quantity.
7. The size-scale decoupling solution simulation system based on spectral physics information neural network according to claim 6 is characterized in that: The size-scale physical quantity decomposition module performs the following operations: is the basis function of the spectral space, and the initial data u(t,x) is expressed as follows in the spectral space basis: Among them, the coefficient U k (t)=(u(t,x),φ k (x)) represents the projection information of the initial data in the spectrum space, x, t represent the spatial coordinate and time coordinate respectively; The main part of the spectral space information, namely the first M spectral information, is retained to obtain large-scale physical quantities.
8. The size-scale decoupling solution simulation system based on spectral physics information neural network according to claim 6 is characterized in that: The input and output of the spectral neural network satisfy the following approximate relationship: in, It is the spectral space information of small-scale physical quantities, neural network Defines the mapping relationship between a single input time t and the small-scale spectral information output, represents a high-dimensional space consisting of positive real numbers and a real vector space of dimension |θ|, is the complementary space of the large-scale spectral space, is a collection of large-scale physical quantity information, represents a preset activation function, and θ represents the set of all parameters to be trained, including weights and offsets.
9. The size-scale decoupling solution simulation system based on spectral physics information neural network according to claim 8 is characterized in that: The loss function used by the spectral neural network is as follows: in, is the loss function of the initial value, which is obtained by projecting the initial value into the spectral space and separating the initial value of the small-scale spectral information; is the partial differential equation loss function satisfied by the small-scale information derived from the physical equation; w PDE and w IC are the weights of the partial differential equation loss function and the initial value loss function respectively; and denote the spectral space transform and inverse transform respectively; Represents nonlinear partial differential operators in partial differential equations; represents the L2 norm of the error, and They represent the true initial conditions and the initial conditions of the spectral neural network output under the parameter θ respectively.
10. The size-scale decoupling solution simulation system based on spectral physics information neural network according to claim 6, characterized in that: The small-scale physical quantity and the large-scale physical quantity in the spectral space output by the spectral neural network after the fusion training are used to obtain the full-scale physical quantity in the spectral space: in, is a large-scale physical quantity, is the small-scale physical quantity output by the spectral neural network under the parameter θ.
Citation Information
Cited By
Dynamic calculation method based on spectral feature coding network
CN120974954A
Power calculation method based on spectral feature coding network
CN120974954B
Nonlinear decoupling method and system of three-axis force sensor
CN120992093A
A method and system for nonlinear decoupling of a triaxial force sensor
CN120992093B
