Combustion simulation method based on physical information neural network coupled small flame model

Through the physical information neural network coupled with the small flame/process variable model, the problem that PINNs cannot consider the detailed reaction mechanism and training difficulty in combustion physics simulation is solved, and the consideration of detailed mechanisms and the acquisition of intermediate components are realized, reducing modeling Cost and training difficulty.

CN118522378BActive Publication Date: 2024-11-12ZHEJIANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410995275.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-12
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

Existing physical information neural networks (PINNs) are difficult to consider detailed chemical reaction mechanisms in combustion physics simulations, and are difficult to train, so it is impossible to effectively obtain intermediate components.

Method used

The physical information neural network is used to couple the small flame/process variable model, and the thermochemical state quantity is parameterized into control parameters through the small flame library, reducing the number of PINNs learning equations, and using the control parameters obtained by PINNs training to quickly obtain the temperature and component distribution.

Benefits of technology

The detailed mechanism consideration of the combustion process is achieved, the difficulty of model training is reduced, the modeling cost is significantly reduced, the intermediate components can be accurately obtained, and the accuracy and efficiency of the simulation are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118522378B_ABST
    Figure CN118522378B_ABST
Patent Text Reader

Abstract

The present invention discloses a combustion simulation method based on a physical information neural network coupled small flame model. The present invention can fully consider the detailed chemical reaction mechanism through the coupling of PINNs and small flame / process variable models; in this process, PINNs only need to learn the basic N-S equation and the equation of the small flame library control parameter, and can efficiently obtain the thermochemical state quantity and thermophysical property parameters such as temperature, component and viscosity / diffusion rate by searching in the small flame library through the control parameters, without solving the temperature / component transport equation and complex ODE calculation, reducing the number of PINNs learning equations and significantly reducing the difficulty of model training; in addition, the small flame library control parameters obtained through PINNs training can quickly obtain the full-field distribution of temperature and different components through post-processing interpolation, which can accurately and effectively solve the deficiency that the current PINNs cannot consider detailed mechanisms in combustion physics problems.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a combustion physics simulation method in the combustion field, and in particular to a combustion physics simulation method of a physical information neural network coupled small flame / process variable model (PINNs / FPV) considering detailed reaction mechanism. Background Art

[0002] Combustion is a complex chemical reaction process that plays a vital role in various fields such as aerospace, fossil energy utilization and thermal power generation. Numerical simulation methods have become a powerful tool for understanding combustion physics. However, combustion problems involve complex reaction chemistry, turbulence and their interactions. Many advances have been made in studying combustion problems using traditional numerical simulation methods, but they rely heavily on high-fidelity models. When considering detailed reaction mechanisms and performing parameterized calculations, the computational and time costs are huge.

[0003] In recent years, with the improvement of computer performance and the advancement of deep learning algorithms, deep learning frameworks that support automatic differentiation (AD) (such as Pytorch, TensorFlow, etc.) have emerged. The automatic differentiation algorithms in these frameworks use the chain rule to realize automatic training of neural networks, making deep neural networks a new method for solving partial differential equations. Physical information neural networks (PINNs) are a machine learning method that uses the powerful approximation ability of neural networks to update network weights through data sampling and backpropagation to obtain approximate solutions to partial differential equations. PINNs can also easily handle inverse problems and avoid the possible data generation costs and grid independence problems in pure data-driven methods. In addition, PINNs show obvious advantages in parameterized simulations. The so-called parameterization refers to the use of PINNs with a series of input parameters to solve the control equations. In traditional computational fluid dynamics (CFD) methods, the control equations are solved for one input parameter value, requiring a large number of simulations to cover a series of input parameters, but parameterized PINNs can obtain a complete solution in a single training.

[0004] Therefore, although PINNs have made significant progress in flow and heat transfer problems, their extension to combustion physics involving strong turbulence, high temperature, mass transfer and their interactions is quite limited, and almost all existing studies use a general package reaction mechanism. The use of purely physically driven PINNs with detailed chemical reaction mechanisms is extremely challenging, and there are currently no reports on such literature. This is mainly because the direct application of PINNs to detailed chemical simulations requires learning a large number of component control equations, which leads to high training difficulty.

[0005] The flamelet / process variable model (FPV) tabulates the thermochemical state quantities using several control parameters and considers turbulent chemical interactions using a pre-assumed probability density function to consider detailed chemical reaction mechanisms at an acceptable computational cost. It assumes that combustion occurs primarily in a thin layer embedded in turbulence, and since the time scale of the chemical reaction is short relative to the time scale of the turbulence, it can be considered as a one-dimensional flame structure. Therefore, the distribution of thermochemical state quantities in the one-dimensional flame structure can be determined by solving the flamelet equation in the mixture fraction space. In actual simulations, a series of flamelet equations reflecting the flame characteristics under different combustion states are solved using detailed chemical reaction mechanisms, and then the solutions to the flamelet equations are tabulated with several control parameters. In CFD simulations, only the transport equations for these control parameters need to be solved.

[0006] Based on the above background, several significant defects of the current PINNs method are found: first, there is a relative lack of detailed research on combustion physics; second, most existing studies use a simplified overall reaction mechanism; third, current research has not fully considered the intermediate components in the combustion process. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the prior art and propose a combustion physics simulation method that considers detailed reaction mechanisms by coupling a physical information neural network with a small flame / process variable model. The present invention can fully consider detailed chemical reaction mechanisms by coupling PINNs with small flame / process variable models; in this process, PINNs only need to learn basic NS equations and equations of small flame library control parameters, and can efficiently obtain thermochemical state quantities and thermophysical parameters such as temperature, components and viscosity / diffusion rate by retrieving control parameters from the small flame library, without having to solve temperature / component transport equations and complex ODE calculations, thereby reducing the number of PINNs learning equations and significantly reducing the difficulty of model training; in addition, the control parameters obtained through PINNs training can be quickly obtained by post-processing interpolation of the full-field distribution of temperature and different components, which can accurately and effectively solve the deficiency of PINNs in combustion physics problems that cannot consider detailed mechanisms.

[0008] The technical solution of the present invention is as follows:

[0009] 1. A combustion simulation method based on physical information neural network coupled small flame model

[0010] 1) Construct a small flame library based on the selected reaction mechanism;

[0011] 2) Determine the geometric structure parameters and boundary conditions according to the target area, and then build a physical information neural network and a loss function of the control equation and boundary conditions based on the geometric structure parameters and boundary conditions;

[0012] 3) Combined with the loss function of the control equation and boundary conditions, the physical information neural network is trained. During the training process, the control parameters of the small flame library based on the network output are extracted from the small flame library to update the thermophysical parameters and process variable source items used in the control equation until the loss function converges to obtain the physical information neural network coupled small flame / process variable model;

[0013] 4) The flow field distribution in the target area and the distribution of the small flame library control parameters are obtained after processing based on the physical information neural network coupled small flame / process variable model.

[0014] In 2), the control equation includes the mixture fraction of the small flame bank Z The control equations and process variables C The control equations of .

[0015] The control equations also include a mass conservation equation and a momentum conservation equation.

[0016] The output of the physical information neural network includes the flow field x Direction Speed u , y Direction Speed v ,pressure p , small flame library control parameters.

[0017] The flamelet library control parameters include the mixture fraction Z and process variables C .

[0018] The loss function of the control equation and the boundary condition is specifically the mean square error between the predicted value and the theoretical value in the control equation and the mean square error between the predicted value of the boundary condition and the set value of the boundary condition.

[0019] The specific contents of 4) are:

[0020] After deriving the physical information neural network coupled small flame / process variable model, the flow field distribution in the target area and the distribution of the small flame library control parameters are obtained; and the temperature field distribution and component field distribution in the target area are obtained by interpolating the distribution of the predicted small flame library control parameters.

[0021] 2. A computer device

[0022] The computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method when executing the computer program.

[0023] 3. A computer-readable storage medium

[0024] The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method are implemented.

[0025] IV. A COMPUTER PROGRAM PRODUCT

[0026] The computer program product comprises a computer program / instructions which, when executed by a processor, implement the steps of the method.

[0027] The beneficial effects of the present invention are as follows:

[0028] 1. Consideration of detailed reaction mechanism: The present invention fully considers the detailed mechanism of the combustion process, can accurately represent the complex reactions in the combustion process, and can obtain intermediate components.

[0029] 2. Low training difficulty: The present invention uses a small flame library to simplify all thermochemical state quantities into control parameter functions, which greatly reduces the number of component equations solved by PINNs and reduces the training difficulty.

[0030] 3. Low modeling cost: The present invention only needs to determine the system design parameters and the definition of the physical problem, and then build a model through a physical information neural network. After the training is completed, the physical quantity parameters of the entire system can be obtained; in addition, the present invention does not rely on any data labels, which significantly reduces the modeling cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the implementation cases of the present invention or the technical solutions in the prior art, the drawings required for use in the implementation cases or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some implementation cases of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0032] Figure 1 The present invention is a flow chart of the method.

[0033] Figure 2 Schematic diagram of the target area in an embodiment of the present invention.

[0034] Figure 3 The cloud diagram of the transport variable results of the embodiment of the present invention; wherein (a) is the velocity of the CFD result u Cloud diagram, (b) is the velocity of CFD results v Cloud diagram, (c) is the mixture fraction of CFD results Z Cloud diagram, (d) is the process variable of CFD results C Cloud map, (e) is the prediction of PINNs x Direction Speed u Cloud map, (f) is the prediction of PINNsy Direction Speed v Cloud map, (g) is the mixed fraction predicted by PINNs Z Cloud diagram, (h) is the process variable predicted by PINNs C Cloud map.

[0035] Figure 4 The transport variable of the embodiment of the present invention is y Quantitative comparison of the distribution of the entire field in the direction; where (a) is the flow field x Direction Speed u Comparison of PINNs prediction value and CFD result, (b) is the flow field y Direction Speed v Comparison of PINNs predictions and CFD results, (c) is the mixture fraction Z Comparison of PINNs prediction values ​​and CFD results, (d) is the process variable C Comparison chart of PINNs prediction values ​​and CFD results.

[0036] Figure 5 The temperature and different components of the embodiment of the present invention are y Quantitative comparison of the full-field distribution of the direction; where (a) is the temperature T Comparison of PINNs predictions and CFD results, (b) is the mass fraction of carbon dioxide Y CO2 Comparison of PINNs prediction value and CFD result, (c) is the carbon monoxide mass fraction Y CO Comparison of PINNs prediction value and CFD result, (d) is the water mass fraction Y H2O Comparison of PINNs prediction value and CFD result, (e) is the hydrogen mass fraction Y H2 Comparison chart of PINNs prediction values ​​and CFD results.

[0037] Figure 6 It is a quantitative comparison diagram of temperature and different components at different positions of an embodiment of the present invention.

[0038] Figure 7 Schematic diagram of the principle of physical information neural network coupling small flame / process variable model. DETAILED DESCRIPTION

[0039] In order to enable those skilled in the art to better understand the present invention, and thus to more clearly define the scope of protection claimed in the present invention, the present invention is described in detail with respect to certain specific embodiments and drawings of the present invention. It should be noted that the following are only certain specific implementation methods of the present invention, which are only part of the embodiments of the present invention, wherein the specific and direct description of the relevant structures is only for the convenience of understanding the present invention, and each specific feature does not naturally and directly limit the scope of implementation of the present invention. The conventional selections and substitutions made by those skilled in the art under the guidance of the present invention, as well as the reasonable arrangement and combination of several technical features under the guidance of the present invention, should all be deemed to be within the scope of protection claimed in the present invention.

[0040] The present invention proposes a combustion physics simulation method of a physical information neural network coupled with a small flame / process variable model that takes into account the detailed reaction mechanism. The method can solve the problem that the detailed reaction mechanism cannot be considered in the existing PINNs for solving combustion problems; it can also solve the problem that a large number of component equations need to be learned when considering detailed chemical simulation in the existing PINNs for solving combustion problems, resulting in high training difficulty; it can also solve the problem that the intermediate components cannot be obtained in the existing PINNs for solving combustion problems. Example

[0041] The core of the present invention is to achieve the goal by coupling the flamelet / process variable model with the physical information neural network (PINNs). For the specific case of two-dimensional laminar mixing layer diffusion flame, the verification process of the PINNs coupling flamelet / process variable model method involves the following key means: (1) Benchmark verification preparation: using the traditional numerical simulation method of OpenFOAM to simulate the flow field and combustion field distribution of the two-dimensional laminar mixing layer diffusion flame, providing a reliable benchmark for the prediction results of PINNs, thereby verifying the accuracy and reliability of the PINNs method; (2) Flamelet library construction: with the help of the flamelet / process variable model table building method, the thermophysical parameters and thermochemical state quantities are completely parameterized as functions of the control parameters, so that PINNs can quickly obtain the required thermophysical parameters through interpolation during the training process, and obtain the thermochemical state quantities in the post-processing interpolation; (3) Neural network architecture design: select an appropriate neural network architecture, take the spatial coordinates as input, and output the velocity field, pressure field, mixture fraction field and process variable field. The network is constructed through the pytorch library, and after hyperparameter debugging, the most suitable neural network architecture for this case is determined; (4) Automatic differentiation method: The automatic differentiation function in the neural network can calculate the partial derivatives of the physical quantities in the region with respect to space, and integrate the residuals of the control equations (partial differential equations) and boundary conditions representing the physical laws into the loss function of the neural network; (5) Back propagation and optimization algorithm: Based on the current parameter values, the parameters are updated in the opposite direction of the gradient (i.e., back propagation), and appropriate optimization methods (such as gradient descent) are used to minimize the loss function until the model reaches the optimal state. (6) Forward reasoning and model evaluation: Based on the trained PINNs model, the parameters are set according to the requirements, and the corresponding prediction results of the flow field and the small flame library control parameters are obtained through the input data. Subsequently, the performance of the model is further evaluated by performance parameters such as relative error to ensure its effectiveness in practical applications.

[0042] The method proposed in the present invention is described below with respect to the problem of two-dimensional laminar mixing layer diffusion flame.

[0043] like Figure 1 As shown, a combustion simulation method based on a physical information neural network coupled small flame / process variable model proposed by the present invention comprises the following steps:

[0044] 1) Construct a small flame library based on the GRI3.0 mechanism; use the small flame / process variable model to use a detailed chemical reaction mechanism to solve a series of small flame equations that reflect the flame characteristics under different combustion conditions, and then tabulate the solutions of the small flame equations with several control parameters to form a small flame library. In the subsequent neural network training, only the transport equations of these control parameters need to be solved. Thermochemical state quantities and thermophysical parameters such as temperature, composition, and viscosity / diffusivity can be interpolated from the small flame library without solving the temperature / composition transport equations.

[0045] 2) Determine the target area, such as Figure 2 As shown, the target area is a rectangle with length x=60 mm and height y=90 mm. The domain is divided into two parts at the left inlet, namely the jet and the accompanying flow. The jet side with a height of 15 mm is injected with pure methane, and the jet side with a height of 15 mm is one of the velocity inlet boundaries (U0=1m / s, Z0=1, C0=0, U0 is the jet inlet velocity, Z0 is the jet inlet mixing fraction, and C0 is the jet inlet process variable), while the remaining 75 mm part is used as the accompanying flow, which is also one of the velocity inlet boundaries (U1=0.1m / s, Z1=0.02, C1=0.09, U1 is the accompanying flow inlet velocity, Z1 is the accompanying flow inlet mixing fraction, and C1 is the accompanying flow inlet process variable), which injects the hot burnout mixture to support fuel combustion. The upper and right sides are both zero pressure outlet boundaries (i.e. p0=0pa, p0 is the pressure of the upper and right outlet boundaries). The lower side is a symmetrical boundary, where the two fluids fully mix, interact and react to form a roughly Figure 2 in the reaction zone.

[0046] The geometric structure parameters and boundary conditions are determined according to the target area, and then the physical information neural network and the loss function of the control equation and boundary conditions are constructed based on the geometric structure parameters and boundary conditions; wherein the boundary conditions include the boundary conditions of the inlet and outlet. In addition to the geometric structure parameters, the physical properties parameters and physical field variables of the fluid also need to be determined. The physical properties of the fluid include the density, viscosity and diffusivity of the fluid; the physical field variables include velocity, pressure, mixture fraction, process variables and their corresponding spatial partial derivatives; the present invention mainly solves the problem that the current PINNs cannot consider the detailed mechanism when solving combustion physics and the training is difficult. For the present invention, that is, the PINNs coupled small flame / process variable model, the control equation includes the mixture fraction of the small flame library Z The control equations and process variables C The governing equations, mass conservation equations, and momentum conservation equations are expressed as follows:

[0047]

[0048] Among them, the flow fieldx Direction Speed u , y Direction Speed v ,pressure p , mixed fraction Z , process variable C They are all solution variables about spatial coordinates. They are used as the output of the neural network, and their partial derivatives can be derived through automatic differentiation function. ρ and μ They represent the density and dynamic viscosity of the gas respectively. D Z and D C is a mixed fraction Z and process variables C The diffusion coefficient of , and based on the assumption of a uniform Lewis number D Z = D C , is the process variable source term.

[0049] The above four thermophysical parameters and process variable source terms can be directly interpolated from the small flame library through the mixture fraction and process variables. Z and process variables C is defined as follows:

[0050] Z = M f / ( M f + M o )

[0051] C=Y CO2 +Y CO +Y H2O +Y H2

[0052] in, M f and M o denote the mass flow of fuel and oxidant in the flow field respectively. The process variable is defined as CO 2 、 CO、H 2 O, H 2 The sum of the mass fractions of Y CO2 、YCO 、Y H2O 、Y H2 They are CO 2 , CO, H 2 O, H 2 For the mixture fraction Z and process variables C , the resolution of the flamelet library is 101×64.

[0053] Based on the geometric structure parameters and boundary conditions, the physical information neural network and the loss function of the control equation and boundary conditions are constructed. Specifically, based on the setting of the physical problem, the network architecture, sampling strategy, optimization algorithm, and hardware resources are designed. Subsequently, a physical-driven neural network model is constructed according to the complexity of the system, and the automatic differentiation function of the neural network is used to calculate the partial derivatives of the variables in space, and the loss function of the specific problem (control equation, boundary condition) is constructed to constrain the training of the neural network.

[0054] 3) Combined with the loss function of the control equation and boundary conditions, the physical information neural network is trained. During the training process, the control parameters of the small flame library based on the network output are extracted from the small flame library to update the thermophysical parameters and process variable source items used in the control equation until the loss function converges to obtain the physical information neural network coupled small flame / process variable model;

[0055] Among them, the output of the physical information neural network includes the flow field x Direction Speed u , y Direction Speed v ,pressure p , small flame library control parameters, small flame library control parameters include mixture fraction Z and process variables C . The loss function of the control equation and boundary conditions is specifically the mean square error between the predicted value and the theoretical value in the control equation and the mean square error between the predicted value of the boundary condition and the set value of the boundary condition. Specifically, the residual of the physical equation is evaluated by moving all terms from the right side of the equation to the left side, and the partial derivatives of the variables can be efficiently calculated using automatic differentiation. After that, the loss is calculated by squaring and averaging the residuals of the sampling points in the region. For example, the loss corresponding to the equation of the process variable ζ pde C It can be expressed as:

[0056]

[0057] in,N r is the number of sampling points in the training domain. This calculation method can be regarded as equivalent to calculating the mean square error between the residual of the equation and 0. MSE ). Therefore, the sum of the losses of all governing equations is ζ pde It can be expressed as:

[0058] ζ pde =ζ pde con +ζ pde moment-x +ζ pde moment-y +ζ pde Z +ζ pde C

[0059] in, MSE( ) is the mean square error; ζ pde con is the loss from the mass conservation equation; ζ pde moment-x is the loss of momentum conservation equation in the x direction; ζ pde moment-y is the loss of momentum conservation equation in the y direction; ζ pde Z is a mixed fraction Z The loss of the control equations; ζ pde C is a process variable C The loss of the control equation.

[0060] For the sampling points on the boundary, in order to ensure that the predicted value on the boundary is infinitely close to the value imposed on the boundary conditions such as the inlet and outlet, the mean square error between the predicted value (indicated by the superscript "pred" and the subscript "bc") and the data of the given boundary conditions (without superscript and indicated by the subscript "bc") is calculated, such as the loss function of the process variable of the inlet condition ζ bc C It can be written as:

[0061]

[0062] in, Nbc is the number of sampling points on the boundary, C bc pred is the process variable on the boundary C The predicted value of C bc is the process variable on the boundary C Then, the mean square error of the residuals of the boundary conditions is ζ bc It can be expressed as:

[0063] ζ bc =MSE(u bc pred -u bc )+MSE(v bc pred -v bc )+

[0064] MSE(p bc pred -p bc )+MSE(Z bc pred -Z bc )+MSE(C bc pred -C bc )

[0065] in, u bc pred It's on the border x Direction Speed u The predicted value of u bc It's on the border x Direction Speed u The setting value of v bc pred It's on the border y Direction Speed v The predicted value of v bc It's on the border yDirection Speed v The setting value of p bc pred is the pressure on the boundary p The predicted value of p bc is the pressure on the boundary p The setting value of Z bc pred is the mixed fraction on the boundary Z The predicted value of Z bc is the mixed fraction on the boundary Z The setting value of

[0066] The training process of the physical information neural network is to accurately calculate the partial derivatives of each physical quantity in the spatial dimension through the automatic differentiation method, and then substitute these partial derivatives into the loss function of the control equation, and comprehensively evaluate the accuracy of the model together with the loss function of the boundary conditions. In each iteration, the predicted small flame library control parameters, namely the mixture fraction and process variables, are used to update the thermophysical parameters of each internal point and the process variable source terms (including density, viscosity / diffusivity and process variable source terms) in real time. These updated parameters are further used to evaluate the loss value to ensure that the loss function can accurately reflect the subtle differences between the model prediction and the actual physical state. Finally, the parameters of the neural network are trained and optimized using back propagation, appropriate optimization algorithms, and learning rate decay strategies to continuously improve the prediction ability and accuracy of the model, and finally construct an accurate physical model designed for a specific system. The model can not only accurately output the values ​​of each physical field variable in the flow field system, but also calculate the partial derivatives of each order of the input parameters, providing strong support for the derivation of the flow field and combustion field distribution in the entire flow field system.

[0067] 4) The combustion condition of the target area is obtained according to the final physical information neural network coupled small flame / process variable model, wherein the combustion condition of the target area is sent to a display device for display to realize the simulation of the combustion condition.

[0068] 4) Specifically:

[0069] The physical information neural network model is used to perform forward reasoning on the specific working conditions in the area to obtain the flow field in the target area and the distribution of the small flame library control parameters; and the temperature field and component field in the target area are obtained by point-by-point interpolation from the small flame library using the predicted distribution of the small flame library control parameters. Figure 3 and Figure 4The qualitative and quantitative comparison between the predicted results of the four variables of PINNs transport and the calculated results of the traditional CFD method is shown in Figure 2. Figure 3 (a) is the velocity of the CFD result u Cloud Atlas, Figure 3 (b) is the velocity of the CFD result v Cloud Atlas, Figure 3 (c) is the mixture fraction of the CFD result Z Cloud Atlas, Figure 3 (d) is the process variable of the CFD result C Cloud Atlas, Figure 3 (e) is the speed predicted by PINNs u Cloud Atlas, Figure 3 (f) is the speed predicted by PINNs v Cloud Atlas, Figure 3 (g) is the mixed fraction predicted by PINNs Z Cloud Atlas, Figure 3 (h) is the process variable predicted by PINNs C Cloud Atlas, Figure 4 (a) is the flow field x Direction Speed u Comparison chart of PINNs prediction value and CFD result, Figure 4 (b) is the flow field y Direction Speed v Comparison chart of PINNs prediction value and CFD result, Figure 4 (c) is a mixed fraction Z Comparison chart of PINNs prediction value and CFD result, Figure 4 (d) is the process variable C Comparison of PINNs predictions and CFD results. Overall, the predictions from the PINN / FPV model are in good agreement with the traditional CFD results obtained using OpenFOAM. Specifically, for the vector field, the PINN-coupled FPV model can well reproduce the jet decay and the low velocity region within the reaction zone caused by the increased viscosity due to combustion. Mixture fraction Z The transport of (passive scalars) can also be captured well. In addition, the process variables C is an active scalar, and its prediction is very consistent with the CFD results. Next, the interpolated temperature field and component field are quantitatively compared with the CFD results. It can be seen intuitively that PINNs reproduces the CFD results very well, such as Figure 5 and Figure 6 As shown, Figure 5 (a) is the temperature T Comparison chart of PINNs prediction value and CFD result, Figure 5 (b) is the mass fraction of carbon dioxide YCO2 Comparison chart of PINNs prediction value and CFD result, Figure 5 (c) is the mass fraction of carbon monoxide Y CO Comparison chart of PINNs prediction value and CFD result, Figure 5 (d) is the water mass fraction Y H2O Comparison chart of PINNs prediction value and CFD result, Figure 5 (e) is the mass fraction of hydrogen Y H2 Comparison chart of PINNs prediction value and CFD result. These results can prove the accuracy of the model, and by analyzing key indicators such as relative error of physical parameters, determination coefficient, normalized root mean square error, etc., the performance of the model can be evaluated from all aspects and angles to ensure its reliability and practicality in complex physical field analysis.

[0070] After completing the quantitative comparison of the performance of the PINNs coupled flamelet / process variable model, further in-depth quantitative statistics were performed to comprehensively evaluate its overall performance. Three performance parameters were used, including mean relative error, MRE ), coefficient of determination, R 2 ) and root mean square error (RMSE, RMSE ), in addition, the velocity component u and v are combined into a composite velocity | u |. Therefore, a total of 8 key parameters were considered, including the combined speed | u | 、 Mixed fractions Z , process variables C、 temperature T.CO 2 Quality score Y CO2 , CO Quality score Y CO 、H 2 O Quality score Y H2O and H 2 Quality score Y H2 In order to more intuitively compare the eight parameters involved, RMSE After normalization, the normalized RMSE ( Normalized-RMSE ),Right now NRMSE This ensures that the parameters are compared at the same order of magnitude. The above individual parameter expressions are defined as follows:

[0071]

[0072] in, x i,CFD represents the CFD results, x i,PINNs represents the predicted value of PINNs, x i,CFD ’ represents the average value of CFD results, x max,CFD and x min,CFD represent the maximum and minimum values ​​in the CFD results respectively, and | | means taking the absolute value.

[0073] Table 1 shows the eight key parameters MRE , R 2 and NRMSE , you can see all the parameters MRE All are less than 10%, the combined speed | u | 、 Mixed fractions Z ,temperature T.CO 2 Quality score Y CO2 and H 2 O Quality score Y H2O Even less than 5%. At the same time, all parameters R 2 All above 0.94, combined speed | u | 、 Mixed fractions Z ,temperature T.CO 2 Quality score Y CO2 and H 2 O Quality score Y H2O Even reaching above 0.99, which shows a very strong correlation. In addition, all parameters NRMSE All of them are lower than 0.07. In summary, the above results show that the PINNs coupled FPV model has comparable ability to the traditional CFD method in predicting velocity and thermochemical state quantities.

[0074] Table 1 shows the performance parameters of 3 of the 8 key parameters.

[0075]

[0076] Figure 7 The schematic diagram of the PINNs coupled flamelet / process variable model for solving two-dimensional steady-state laminar combustion problems is shown. It consists of two parts. The first part is the flamelet library in the FPV model, and the second part is the PINNs architecture. The input of the model is the spatial coordinates and, and the output of the model includes the variables to be solved ( u, v, p, Z and C ). Before the input and output of the model, several neural network hidden layers are connected to build a nonlinear mapping relationship between spatial coordinates and output variables. Figure 7 As shown, the first five loss terms of the model are the losses of the control equation ζ pde , which reflects the difference between the model prediction and the true physical law. The sixth loss term is the loss of boundary conditions ζ bc , which describes the deviation between the model prediction and the true value at the boundary. During the training process, according to the predicted mixture fraction Z and process variables C , extract the thermophysical parameters from the small flame library (i.e. ρ, µ, D Z 、D C ) and process variable source terms to update these parameters used in the equation. The loss function of the boundary conditions and the loss function of the partial differential equation (PDE) are then aggregated into the total loss function. When the total loss exceeds the preset threshold, the trainable weights in the neural network are continuously adjusted through back propagation and optimization algorithms until the total loss function is reduced to an acceptable range, which means that the training process is completed.

[0077] In summary, the present invention first determines to verify the method in a two-dimensional laminar mixing layer diffusion flame, and clarifies the key parameters such as the geometric structure, boundary conditions, and combustion mechanism required for the study. The flamelet / process variable model uses a detailed chemical reaction mechanism to solve a series of flamelet equations that reflect the flame characteristics under different combustion states. Then the solution of the flamelet equation and several control parameters are used to tabulate the combustion-related thermophysical parameters and thermochemical state quantities; then, a physically driven neural network model is established based on the physical problem. The model uses the automatic differentiation function to calculate the partial derivatives of the physical variables in the space, and is brought into the control equation to calculate the residual, and the thermophysical parameters required by the control equation can be directly obtained by interpolation through the flamelet library. Through back propagation and optimization algorithms, the neural network weights are continuously updated to minimize the loss function; finally, the trained model is used to perform forward reasoning to obtain the flow field distribution and the distribution of the flamelet library control parameters under different conditions, and the temperature field distribution and component field distribution of the target area are retrieved from the flamelet library through the predicted distribution of the flamelet library control parameters, and compared with the results of traditional numerical simulation, and the performance of the model is further analyzed and evaluated. The present invention not only considers the detailed reaction mechanism, but also tabulates the thermochemical state quantity by coupling the small flame / process variable model with PINNs. PINNs only need to solve the control equations of continuity, momentum and control parameters, which greatly reduces the number of PINNs learning equations and reduces the training difficulty. In addition, the physical drive method does not need to collect any data labels in advance, which effectively reduces the modeling cost.

[0078] On the one hand, the present invention proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of a combustion simulation method based on a physical information neural network coupled with a small flame model.

[0079] On the one hand, the present invention proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a combustion simulation method based on a physical information neural network coupled with a small flame model.

[0080] On the one hand, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of a combustion simulation method based on a physical information neural network coupled with a small flame model.

[0081] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A combustion simulation method based on a physical information neural network coupled small flame model, characterized in that: The following steps are involved: 1) Construct a small flame library based on the selected reaction mechanism; 2) Determine the geometric structure parameters and boundary conditions according to the target area, and then build a physical information neural network and a loss function of the control equation and boundary conditions based on the geometric structure parameters and boundary conditions; 3) Combined with the loss function of the control equation and boundary conditions, the physical information neural network is trained. During the training process, the control parameters of the small flame library based on the network output are extracted from the small flame library to update the thermophysical parameters and process variable source items used in the control equation until the loss function converges to obtain the physical information neural network coupled small flame / process variable model; 4) The flow field distribution in the target area and the distribution of the control parameters of the small flame library are obtained after processing based on the physical information neural network coupling small flame / process variable model; In 2), the control equation includes the mixture fraction of the small flame bank Z The control equations and process variables C The control equation is as follows: Among them, the flow field x Direction Speed u , y Direction Speed v ,pressure p , mixed fraction Z , process variable C are all variables to be solved about spatial coordinates, which are used as the output of the neural network. ρ Indicates the density of the gas; D Z and D C is a mixed fraction Z and process variables C The diffusion coefficient of D Z = D C , is the process variable source term; The loss function of the control equation and the boundary condition is specifically the mean square error between the predicted value and the theoretical value in the control equation and the mean square error between the predicted value of the boundary condition and the set value of the boundary condition; The sum of the mean square errors between the predicted values ​​and the theoretical values ​​in all control equations is taken as the sum of the losses of all control equations. ζ pde , expressed as: ζ pde =ζ pde con + ζ pde moment-x +ζ pde moment-y +ζ pde Z +ζ pde C in, ζ pde con is the loss from the mass conservation equation; ζ pde moment-x is the loss of momentum conservation equation in the x direction; ζ pde moment-y is the loss of momentum conservation equation in the y direction; ζ pde Z is a mixed fraction Z The loss of the control equations; ζ pde C is a process variable C The loss of the control equation.

2. The combustion simulation method based on physical information neural network coupled small flame model according to claim 1 is characterized in that: The control equations also include a mass conservation equation and a momentum conservation equation.

3. The combustion simulation method based on physical information neural network coupled small flame model according to claim 1 is characterized in that: The output of the physical information neural network includes the flow field x Direction Speed u , y Direction Speed v ,pressure p , small flame library control parameters.

4. The combustion simulation method based on physical information neural network coupled small flame model according to claim 3 is characterized in that: The flamelet library control parameters include the mixture fraction Z and process variables C .

5. The combustion simulation method based on physical information neural network coupled small flame model according to claim 1 is characterized in that: The specific contents of 4) are: After deriving the physical information neural network coupled small flame / process variable model, the flow field distribution in the target area and the distribution of the small flame library control parameters are obtained; and the temperature field distribution and component field distribution in the target area are obtained by interpolating the distribution of the predicted small flame library control parameters.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.