A numerical calculation method for plasma equations based on a dual neural network framework
Through the method based on the dual neural network framework, the problem of traditional numerical simulation methods relying on meshing and high number of iterations in plasma equation solving is solved, and the precise numerical calculation of plasma equations is realized.
Patent Information
- Application Number
- CN202110782738.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-07-12
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-07-12
AI Technical Summary
Traditional numerical simulation methods such as finite difference, finite element, and finite volume methods rely on meshing when solving plasma equations, resulting in inaccurate accuracy of high-dimensional problems, and high-order solutions require a lot of iteration.
Using a method based on the bi-neural network framework, by establishing a plasma equation model, physical laws are used as prior information of the neural network, loss functions are constructed and neural networks are trained to solve plasma equations, and the calculation of the precise solution is realized.
The corresponding mapping set is effectively trained from finite data samples, solving the problem of mesh dependence and high number of iterations, and realizing the precise numerical calculation of plasma equations.
Smart Images

Figure CN113297534B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of artificial intelligence and plasma technology, and in particular to a plasma equation numerical calculation method based on a dual neural network framework. Background Art
[0002] Plasma is a form of matter with free electrons and charged ions as its main components. It is widely present in the universe and is often regarded as the fourth state of matter, also known as the plasma state. Plasma has high electrical conductivity and is strongly coupled with electromagnetic fields. It is crucial to accurately understand the characteristics of plasma and its development process. Take the electric arc as an example. As a high-temperature, high-conductivity discharge plasma, the electric arc can generate temperatures of tens of thousands of degrees Celsius during discharge, and also generate a large amount of thermal radiation.
[0003] Numerical simulation is a common method for studying plasma characteristics, including finite difference, finite element, and finite volume methods. However, these traditional methods have certain defects. Their results rely on grid division, and there may be inaccurate precision in solving high-dimensional problems. Artificial neural networks are a powerful nonlinear mapping tool with great potential for solving partial differential equations. When using artificial neural networks to solve plasma equations, first construct an artificial neural network to describe the plasma equation model. The prior information of the network is based on the physical laws satisfied by the plasma equation. Then, the corresponding loss functions are designed, the width, depth, and activation function of the neural network are selected, and the computational solution of the plasma partial differential equation is obtained by batch training. Summary of the invention
[0004] The purpose of the present invention is to provide a numerical calculation method for the plasma equation based on a dual neural network framework, which utilizes the dual neural network architecture to solve the exact solution of the plasma equation, thereby solving the defects of the conventional effective volume method that the results depend on grid division and high-order solutions require a large number of iterations, and can effectively train the corresponding mapping set from limited data samples.
[0005] The purpose of the present invention can be achieved through the following technical solutions:
[0006] A numerical calculation method for plasma equations based on a dual neural network framework comprises the following steps:
[0007] S1: Establish a plasma equation model and use the physical laws contained in the plasma equation as the prior information of the artificial neural network;
[0008] S2: Establish the corresponding dual neural network framework based on the plasma equation of S1;
[0009] S3: construct the loss function based on the equation and the corresponding boundary conditions and initial conditions, and select appropriate parameters such as the number of layers of the neural network;
[0010] S4: The neural network is trained to obtain the numerical solution of the plasma equation. New loss function values are continuously obtained during the training. When the loss function converges to the threshold, the training is terminated, thereby realizing the artificial neural network solution of the plasma equation model.
[0011] Furthermore, the S1 specifically includes:
[0012] S11: Establish a corresponding plasma model according to specific problems;
[0013] S12: Rewrite the corresponding plasma model into the following formula:
[0014]
[0015] The boundary conditions and initial conditions are:
[0016] f b (u,t,x,y,z)=0
[0017] f i (u,x,y,z)=0
[0018] Among them, x, y, z are spatial coordinates in three directions, t is the time coordinate, u is the solution of the equation, and the specific meaning depends on the type of plasma equation, and λ is a parameter representing the properties of the plasma.
[0019] Furthermore, the S2 specifically includes:
[0020] S21: according to the plasma equation, the independent variable of the equation is used as the input of the neural network, and the solution of the equation is used as the output value of the neural network;
[0021] S22: embed a small neural network, take some relevant parameters λ of the equation as the output of the small neural network, and calculate the independent variable of λ as the input of the small neural network;
[0022] S23: The output obtained by training the small neural network is used as the value of the corresponding parameter of the plasma equation.
[0023] Furthermore, the S3 specifically includes:
[0024] S31: Select a suitable activation function;
[0025] S32: Construct the first part L of the loss function according to the plasma equation f ;
[0026] S33: Construct the second part L of the loss function according to the boundary conditionsb ;
[0027] S34: Construct the third part L of the loss function based on the initial conditions i ;
[0028] S35: construct loss function L=L f +L b +L i ;
[0029] S36: Select the appropriate number of neural network layers and the number of neurons in each layer.
[0030] Furthermore, the activation function of S31 is the Huber function:
[0031]
[0032] Furthermore, L in S32 f The calculation formula is as follows:
[0033]
[0034] Where N f is the number of sampling points in the computational domain, and Ψ is the huber function.
[0035] Furthermore, L in S33 b The calculation formula is as follows:
[0036]
[0037] Where N b is the number of sampling points within the boundary domain.
[0038] Furthermore, L in S34 i The calculation formula is as follows:
[0039]
[0040] Where N i is the number of sampling points in the boundary domain; if no initial conditions are given, L i =0.
[0041] Furthermore, the S4 specifically includes:
[0042] S41: Observe the loss function value of the neural network until it drops to a given threshold;
[0043] S42: Observe the L2 norm error value of the neural network until it drops to a given threshold. The L2 norm is the distance between two points in the feature space. If there are points A(x1, y1) and B(x2, y2) in the space, the L2 norm error between points A and B is:
[0044]
[0045] S43: Obtain the output of the neural network, i.e., the numerical solution of the corresponding plasma equation.
[0046] Beneficial effects of the present invention:
[0047] The numerical calculation method of the plasma equation of the present invention utilizes a dual neural network architecture to solve the exact solution of the plasma equation, which solves the defects of the conventional effective volume method that the results depend on grid division and high-order solutions require a large number of iterations, and can effectively train the corresponding mapping set from limited data samples. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] The present invention will be further described below in conjunction with the accompanying drawings.
[0049] Figure 1 is a flow chart of the method of the present invention;
[0050] Figure 2 It is a schematic diagram of a dual neural network for solving a 1-dimensional steady-state arc model of the present invention;
[0051] Figure 3 is a graph of the loss function value of the neural network of the present invention;
[0052] Figure 4 is the L2 error graph of the neural network of the present invention;
[0053] Figure 5 It is a comparison diagram of the neural network training results of the present invention and the calculation results using the finite element method;
[0054] Figure 6 It is a comparison diagram of the training results of the neural network of the present invention and the results calculated by the finite element method;
[0055] Figure 7 It is a comparison diagram of the results obtained by the neural training of the present invention and the results calculated by the finite element method. DETAILED DESCRIPTION
[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0057] Example 1
[0058] In this embodiment, a 1D steady-state arc is taken as a research object, SF6 gas is taken as a research sample, and a numerical solution of the 1D steady-state arc equation is obtained by batch training.
[0059] See also Figure 1 , which shows a flowchart of numerical solution of 1-dimensional steady-state arc equation based on dual neural network, the method includes the following steps:
[0060] S1: Establish a general model of 1D steady-state arc equation;
[0061] S11: Establish a 1D arc equation model based on the mass conservation equation, energy conservation equation and Ohm's law equation:
[0062]
[0063]
[0064] When the arc is in a steady state, the speed and time variables are both 0, so the simplified formula is:
[0065]
[0066] S12: Rewrite the corresponding 1D steady-state arc equation model into the following general form:
[0067]
[0068] The boundary conditions are:
[0069] T| r=R =T b
[0070]
[0071] Where ρ is the density, t is the time, r is the arc radius, v r is the arc speed, C p is the specific heat, T is the temperature, σ is the electrical conductivity, g is the arc conductance, k is the thermal conductivity, E rad is the energy loss caused by radiation, T b The boundary temperature value is given when r = R, and the parameters λ representing the plasma properties are: σ, k, E rad .
[0072] S2: Based on the plasma equation of S1, a corresponding dual neural network framework is established, such as Figure 2 As shown;
[0073] S21: according to the plasma equation, the independent variable of the equation is used as the input of the neural network, and the solution of the equation is used as the output value of the neural network;
[0074] S22: embed a small neural network, take some relevant parameters λ of the equation as the output of the small neural network, and calculate the independent variable of λ as the input of the small neural network;
[0075] S23: Substitute the output obtained from the training of the small neural network into the value of the corresponding parameter of the plasma equation.
[0076] S3: Based on the 1D steady-state arc equation of S1, the loss function is constructed based on the equation and the corresponding boundary conditions, and the appropriate number of neural network layers and other parameters are selected;
[0077] S31: Select a suitable activation function, i.e., Huber function;
[0078]
[0079] S32: Construct the first part L of the loss function according to the plasma equation f , the calculation formula is as follows:
[0080]
[0081] Where N f is the number of sampling points in the computational domain;
[0082] S33: Construct the second part L of the loss function according to the boundary conditions b , the calculation formula is as follows:
[0083]
[0084] Where N b is the number of sampling points within the boundary domain;
[0085] S34: Construct the third part L of the loss function based on the initial conditions i , there is no initial condition in the case of a 1D steady-state arc, so L i =0;
[0086] S35: construct loss function L=L f +L b +L i =L f +L b ;
[0087] S36: Set the neural network hidden layer to 5 layers, 50 neurons in each layer, randomly initialize the weights, and set the learning rate to 10 -4 , the number of training times is set to 100000 times.
[0088] S4: The neural network is trained to obtain the numerical solution of the plasma equation, and a new loss function value is continuously obtained during the training. When it converges to a certain threshold, the training is terminated;
[0089] S41: The loss function value obtained changes as follows Figure 3 As shown in the figure, in the second half of the training, the loss function value oscillates. When the loss function value has converged to a certain small range, it will oscillate repeatedly within this range;
[0090] S42: The obtained L2 norm error value changes as follows Figure 4 As shown, the L2 norm is the distance between two points in the feature space. If there are points A(x1,y1) and B(x2,y2) in the space, the L2 norm error between points A and B is:
[0091]
[0092] from Figure 4 It can be seen that in the second half of the training, the L2 norm error value also oscillates. When the L2 norm error value has converged to a certain threshold, it will oscillate repeatedly within this range, ensuring that the network will not overfit;
[0093] S43: Comparison of the obtained neural network training results and the results calculated using the finite element method. Figure 5 As shown. From the comparison between the loss error and the result, under this model, after 100,000 training times, the loss error converged to a very small value, ensuring the convergence of the neural network, and the temperature value obtained by training is also very close to the fitting curve of the temperature value calculated by the finite element method. It can be concluded that the artificial neural network designed in this embodiment can well calculate the one-dimensional steady-state arc fluid equation of SF6 gas and obtain a more accurate numerical solution to the equation.
[0094] Example 2
[0095] In this embodiment, a 1D steady-state arc is used as a research object, Ar gas is used as a research sample, and a numerical solution of the 1D steady-state arc equation is obtained by batch training. The method includes the following steps:
[0096] S1: Establish a general model of 1D steady-state arc equation;
[0097] S11: Establish a 1D arc equation model based on the mass conservation equation, energy conservation equation and Ohm's law equation:
[0098]
[0099]
[0100] When the arc is in a steady state, the speed and time variables are both 0, so the simplified formula is:
[0101]
[0102] S12: Rewrite the corresponding 1D steady-state arc equation model into the following general form:
[0103]
[0104] The boundary conditions are:
[0105] T| r=R =T b
[0106]
[0107] Where ρ is the density, t is the time, r is the arc radius, and v r is the arc speed, C p is the specific heat, T is the temperature, σ is the electrical conductivity, g is the arc conductance, k is the thermal conductivity, E rad is the energy loss caused by radiation, T b The boundary temperature value is given when r = R, and the parameters λ representing the plasma properties are: σ, k, E rad .
[0108] S2: Establish the corresponding dual neural network framework based on the plasma equation of S1;
[0109] S21: according to the plasma equation, the independent variable of the equation is used as the input of the neural network, and the solution of the equation is used as the output value of the neural network;
[0110] S22: embed a small neural network, take some relevant parameters λ of the equation as the output of the small neural network, and calculate the independent variable of λ as the input of the small neural network;
[0111] S23: Substitute the output obtained from the training of the small neural network into the value of the corresponding parameter of the plasma equation.
[0112] S3: Based on the 1D steady-state arc equation of S1, the loss function is constructed based on the equation and the corresponding boundary conditions, and the appropriate number of neural network layers and other parameters are selected;
[0113] S31: Select a suitable activation function, i.e., Huber function;
[0114]
[0115] S32: Construct the first part L of the loss function according to the plasma equation f , the calculation formula is as follows:
[0116]
[0117] Where N fis the number of sampling points in the computational domain;
[0118] S33: Construct the second part L of the loss function according to the boundary conditions b , the calculation formula is as follows:
[0119]
[0120] Where N b is the number of sampling points within the boundary domain;
[0121] S34: Construct the third part L of the loss function based on the initial conditions i , there is no initial condition in the case of a 1D steady-state arc, so L i =0;
[0122] S35: construct loss function L=L f +L b +L i =L f +L b ;
[0123] S36: Set the neural network hidden layer to 6 layers, 50 neurons in each layer, randomly initialize the weights, and set the learning rate to 10 -4 , the number of training times is set to 50000 times.
[0124] S4: The neural network is trained to obtain the numerical solution of the plasma equation. During the training, new loss function values are continuously obtained. When it converges to a certain threshold, the training is terminated. The obtained neural network training results are compared with the results calculated by the finite element method. Figure 6 As shown. From the comparison between the loss error and the result, under this model, after 50,000 trainings, the loss error converged to a very small value, ensuring the convergence of the neural network. The temperature value obtained by training is also very close to the fitting curve of the temperature value calculated by the finite element method. It can be concluded that the artificial neural network designed in this embodiment can well calculate the one-dimensional steady-state arc fluid equation of Ar gas and obtain a more accurate numerical solution to the equation.
[0125] Example 3
[0126] In this embodiment, a 1D transient arc is used as the research object, SF6 gas is used as the research sample, and a numerical solution of the 1D transient arc equation is obtained by batch training. The method includes the following steps:
[0127] S1: Establish a general model of 1D transient arc equation;
[0128] S11: Establish a 1D arc equation model based on the mass conservation equation, energy conservation equation and Ohm's law equation:
[0129]
[0130]
[0131] S12: Rewrite the corresponding 1D steady-state arc equation model into the following general form:
[0132]
[0133] The boundary conditions are:
[0134] T| r=R =T b
[0135]
[0136] The initial conditions are:
[0137] T| t=0 =T i
[0138] Where ρ is the density, t is the time, r is the arc radius, v r is the arc speed, C p is the specific heat, T is the temperature, σ is the electrical conductivity, g is the arc conductance, k is the thermal conductivity, E rad is the energy loss caused by radiation, T b The boundary temperature value is given when r = R, and the parameters λ representing the plasma properties are: σ, k, E rad .
[0139] S2: Establish the corresponding dual neural network framework based on the plasma equation of S1;
[0140] S21: according to the plasma equation, the independent variable of the equation is used as the input of the neural network, and the solution of the equation is used as the output value of the neural network;
[0141] S22: embed a small neural network, take some relevant parameters λ of the equation as the output of the small neural network, and calculate the independent variable of λ as the input of the small neural network;
[0142] S23: Substitute the output obtained from the training of the small neural network into the value of the corresponding parameter of the plasma equation.
[0143] S3: Based on the 1-dimensional transient arc equation of S1, the loss function is constructed based on the equation and the corresponding boundary conditions, and the appropriate number of neural network layers and other parameters are selected;
[0144] S31: Select a suitable activation function, i.e., Huber function;
[0145]
[0146] S32: Construct the first part L of the loss function according to the plasma equation f , the calculation formula is as follows:
[0147]
[0148] Where N f is the number of sampling points in the computational domain;
[0149] S33: Construct the second part L of the loss function according to the boundary conditions b , the calculation formula is as follows:
[0150]
[0151] Where N b is the number of sampling points within the boundary domain;
[0152] S34: Construct the third part L of the loss function based on the initial conditions i , the calculation formula is as follows:
[0153]
[0154] Where N i is the number of sampling points within the boundary domain;
[0155] S35: construct loss function L=L f +L b +L i ;
[0156] S36: Set the neural network hidden layer to 6 layers, 100 neurons in each layer, randomly initialize the weights, and set the learning rate to 10 -4 , the number of training times is set to 100000 times.
[0157] S4: The neural network is trained to solve the numerical solution of the plasma equation. During the training, new loss function values are continuously obtained. When it converges to a certain threshold, the training ends. Take t = 100us, 500us, and 900us respectively, and observe the comparison between the results obtained by neural training and the results calculated by finite element method. Figure 7 As shown. From the comparison between the loss error and the result, under this model, after 100,000 training times, the loss error converged to a very small value, ensuring the convergence of the neural network. The temperature and velocity values obtained by training are also very similar to the fitting curves of the temperature and velocity values calculated by the finite element method. It can be concluded that the artificial neural network designed in this embodiment can well calculate the one-dimensional transient arc fluid equation of SF6 gas and obtain a more accurate numerical solution to the equation.
[0158] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.
[0159] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.
Claims
1. A one-dimensional steady-state arc equation calculation method based on dual neural networks, characterized in that: The steps include: S1: Establish the general model of the one-dimensional steady-state arc equation of SF6 gas and the one-dimensional steady-state arc equation; specifically include: S11: Establish a one-dimensional arc equation model based on the mass conservation equation, energy conservation equation and Ohm's law equation: When the arc is in a steady state, the speed and time variables are both 0, so the general model of the one-dimensional steady-state arc equation can be obtained by simplifying the above formula: S12: Rewrite the corresponding one-dimensional steady-state arc equation general model into the following general form to obtain the one-dimensional steady-state arc equation, i.e., the plasma equation: The boundary conditions are: in, is the density, t It's time. is the arc speed, is the specific heat, T is the temperature, is the conductivity, g is the arc conductance, k is the thermal conductivity, is the energy loss due to radiation, for r=R The boundary temperature value given at ; S2: Based on the plasma equation of S1, a corresponding dual neural network framework is established, including: S21: according to the plasma equation, the independent variable of the equation is used as the input of the neural network, and the solution of the equation is used as the output value of the neural network; S22: Embed a small neural network to represent the parameters of plasma properties As the output of the small neural network, compute The independent variables are used as inputs to a small neural network; is a parameter representing the properties of plasma, including , k , ; S23: Substituting the output obtained by training the small neural network into the value of the corresponding parameter of the plasma equation; S3: Based on the one-dimensional steady-state arc equation of S1, the loss function is constructed based on the equation and the corresponding boundary conditions, and the appropriate number of layers of the dual neural network is selected, including: S31: Select Huber activation function; S32: Constructing the first part of the loss function according to the plasma equation , the calculation formula is as follows: in is the number of sampling points in the computational domain; S33: Construct the second part of the loss function based on boundary conditions , the calculation formula is as follows: in is the number of sampling points within the boundary domain; S34: Construct the third part of the loss function based on the initial conditions , there is no initial condition in the case of a one-dimensional steady-state arc, so L i =0; S35: Construct loss function ; S36: Set the neural network hidden layer to 5 layers, 50 neurons in each layer, randomly initialize the weights, and set the learning rate to 10 -4 , the number of training times is set to 100000 times; S4: dual neural network training to solve the numerical solution of the plasma equation, and continuously obtain new loss function values during training. When it converges to the threshold, the training is terminated, thereby realizing the dual neural network solution of the plasma equation; The S4 specifically includes: S41: observing the loss function value of the dual neural network until it drops to a given threshold; S42: Observation of the dual neural network L 2 norm error value until it drops to a given threshold, L The 2-norm is the distance between two points in the feature space. , ,but A. B Two o'clock L The 2-norm error is: S43: Obtain the output of the dual neural network, which corresponds to the numerical solution of the plasma equation.
