A method for calculating environmental gas chemical kinetics based on neural ordinary differential equation

Abstract

The application discloses an environmental gas chemical kinetics calculation method based on a neural ordinary differential equation, and comprises the following steps: according to a constructed environmental gas chemical reaction kinetics model, a neural ordinary differential equation and a loss function are constructed and reduced to obtain a neural ordinary differential model; network training is performed on the neural ordinary differential model based on a training and test data set, an adaptive loss function weight is adjusted to update a weight coefficient, and a total network loss function is calculated; the network after the weight coefficient is updated is calculated by back propagation to obtain a gradient, a gradient descent method is used to update network parameters until the loss function is reduced to a given threshold value, and the result of the evolution of the concentration of each particle in the reaction system with time is obtained. The application can more stably and efficiently perform model simulation calculation in a strong rigid environmental gas chemical reaction system, can balance the gradient of different tasks, improve the learning ability of the model to the multi-task process, and significantly improve the calculation precision.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and environmental gases, specifically involving a method for calculating the chemical kinetics of environmental gases based on neural ordinary differential equations. Background Technology

[0002] With the development of ultra-high voltage power transmission technology, gas-insulated equipment has been widely used in the power industry. Traditional SF6 insulating gas, due to its strong greenhouse effect and increasingly stringent environmental constraints, is gradually facing usage restrictions. To meet the insulation requirements of power equipment, new environmentally friendly insulating gases are gradually becoming alternatives to SF6. However, in actual operation, due to factors such as insulation defects, partial discharge can easily occur inside gas-insulated equipment, leading to the decomposition of the environmentally friendly gas. The resulting reaction products may cause insulation performance degradation, thus threatening the safe and stable operation of the power system. Therefore, it is necessary to perform high-precision modeling and calculation of the evolution of particle concentration over time in the chemical kinetics of environmentally friendly gases to accurately reveal its decomposition mechanism and insulation degradation process.

[0003] Existing numerical methods for handling such chemical kinetic problems are often hampered by the multi-scale coupling, high nonlinearity, and rigidity of the system. Traditional numerical methods require extremely small step sizes to ensure stability under rigid conditions, leading to excessive computational resource consumption and potentially limiting the accuracy of results due to accumulated errors. While conventional deep learning-based neural network differential equation methods can model chemical equations by learning the source terms, they are prone to gradient imbalances and training instability in rigid chemical systems, making it difficult to obtain high-precision solutions.

[0004] Therefore, there is an urgent need to propose an improved method for modeling and calculating neural ordinary differential equations to break through the accuracy bottleneck of traditional numerical integration methods and conventional neural network methods, so as to achieve highly reliable simulation and calculation of environmental gas chemical dynamics processes and provide more accurate data support for the insulation performance evaluation and operational safety of power equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a method for calculating the chemical dynamics of environmentally friendly gases based on the neural ordinary differential equation, which can perform model simulation calculations more stably and efficiently in the rigid environmentally friendly gas chemical reaction system. Compared with traditional methods, this invention can balance the gradients of different tasks, improve the model's learning ability for multi-task processes, and significantly improve the calculation accuracy.

[0006] To achieve the above objectives, the solution of the present invention is:

[0007] A method for calculating the chemical kinetics of environmentally friendly gases based on the constant differential equations of God includes the following steps:

[0008] Step 1: Based on the feasible reaction pathways and reaction particles of the environmentally friendly gas reaction system, construct a chemical reaction kinetic model of the environmentally friendly gas, and construct a set of ordinary differential equations of chemical kinetics based on the environmentally friendly gas chemical reaction kinetic model;

[0009] Step 2: Based on the chemical kinetic ordinary differential equations, construct the neural ordinary differential equations and loss function, and reduce the neural ordinary differential equations and loss function to obtain the neural ordinary differential model.

[0010] Step 3: Train the neural network based on the training and testing datasets, and use an adaptive loss function to adjust and update the weight coefficients, and calculate the total network loss function;

[0011] Step 4: Calculate the gradient of the updated weight coefficients network through backpropagation, and update the network parameters using the gradient descent method until the loss function decreases to a given threshold, thus obtaining the result of the evolution of the concentration of each particle in the reaction system over time.

[0012] The specific process of step 1 above is as follows:

[0013] The feasible reaction pathways of the environmentally friendly gas reaction system are determined using quantum chemistry and transition state theory. The reaction rate constants for each reaction at a given temperature are calculated, and the reaction particles are identified.

[0014] The following chemical kinetic model is constructed to describe a chemical reaction involving M types of reacting particles and K types of reactions.

[0015] ,

[0016] in, Let i be the chemical symbol for the i-th particle. and These are the stoichiometric coefficients of the reactants and products of the i-th particle in the k-th reaction, respectively.

[0017] A set of ordinary differential equations is constructed to describe the evolution of the particles participating in the reaction over time under a given initial reaction condition.

[0018] ,

[0019] in, Let be the concentration of the i-th type of particle. Let i be the stoichiometric coefficient of the i-th particle in the k-th reaction. Let be the reaction rate of the k-th reaction.

[0020] In step 2 above, based on the aforementioned set of chemical kinetic ordinary differential equations, the neural network's ordinary differential equations are constructed as follows:

[0021] ,

[0022] Where Y is the particle concentration vector, and t is time. These are network parameters; a neural network (NN) consists of parameters. Defined neural network function;

[0023] Construct the loss function of the neural network based on the training task.

[0024] ,

[0025] Where N is the number of data pairs, loss is the error calculation function, and Y* is the predicted particle concentration vector value. It is the starting time of the i-th time segment. It is the end time of the i-th time segment. Arrehenius is the source term of the chemical kinetic model obtained based on the Arrhenius formula, and NN is the source term function of the chemical kinetic model obtained based on the neural network.

[0026] By reducing the constant differential equation of God, we obtain:

[0027] ,

[0028] Among them, Y scale Y is the vector of maximum particle concentrations across all particles in the dataset. max Subtract the minimum value vector Y min The particle concentration reduced scale vector, t scale It is the time-reduced scale vector obtained by subtracting the minimum value t0 from the maximum value t1 of all times in the dataset;

[0029] After reducing the loss function, we get:

[0030] .

[0031] In step 3 above, the initial and final values ​​of the concentration of reactive particles over a given time period are obtained using the numerical method of the Arrhenius formula, forming the training and testing datasets.

[0032] Step 3 above specifically includes the following steps:

[0033] Step 31: Train the God frequent differential model and initialize the task weight parameters corresponding to different reaction particles. Calculate the initial loss function ;

[0034] Step 32: At the t-th iteration of the network, calculate the loss function corresponding to different reaction particles. And calculate the loss function for each particle concentration relative to the network parameters. gradient,

[0035] ,

[0036] Then calculate its L2 norm.

[0037] ,

[0038] Define gradient norm As the average baseline gradient norm that all tasks hope to achieve

[0039] ,

[0040] Calculate the convergence rate of the i-th particle. ,

[0041] ,

[0042] Then calculate the relative convergence velocity of the i-th particle. ,

[0043] ,

[0044] The target norm for obtaining the gradient norm of each particle learning task ,

[0045] ,

[0046] in, It is the convergence rate adjustment factor, and its value range is... ;

[0047] Step 33, obtain information about the weights. loss function,

[0048] ,

[0049] Differentiating the loss function yields The corresponding gradient, thus affecting Update;

[0050] Step 34: The total loss function of the network is calculated by weighting and summing the loss functions of different particles according to the weight coefficients.

[0051] In step 3 above, the updating of the loss function weight coefficients is decoupled from the updating of network parameters in parallel. The loss function weights are updated multiple times during a single network parameter iteration update or after multiple network parameter iteration updates.

[0052] Step 4 above specifically includes the following steps:

[0053] Step 41, determine the loss L for the hidden state at each time step. The gradient dependency, i.e., the associated state Its dynamic changes are described by the following ordinary differential equation:

[0054] ,

[0055] Calculate relative to network parameters gradient,

[0056] ,

[0057] Step 42, use all the solutions , as well as The integrals are synthesized into an extended system of ordinary differential equations, which... Treating both and t as variables in the system of ordinary differential equations, we obtain the following equation:

[0058] ,

[0059] On this basis and By combining these, we obtain augmented variables that have a system of ordinary differential equations for the response and the adjoint state.

[0060] ,

[0061] ,

[0062] ,

[0063] ,

[0064] Step 43: Solve this augmented equation to obtain the loss function L with respect to the network parameters. The gradient is used to update the network parameters using gradient descent.

[0065] Step 44: Repeat steps 41-43 until the loss function drops to a given threshold;

[0066] Step 45 yields the network output, which is the result of the evolution of the concentration of each particle in the reaction system over time.

[0067] After adopting the above solution, the beneficial effects of the present invention compared with the prior art are as follows:

[0068] (1) In the face of the complex environmental gas chemical dynamics modeling and calculation problem with strong rigidity and multi-scale coupling, the method of the present invention transforms the discrete stepping process in the traditional method into a continuous dynamic evolution process, and reduces the computational dimension and gradient storage overhead through reduction processing and backpropagation gradient calculation, thereby significantly improving computational efficiency and stability while maintaining accuracy.

[0069] (2) Compared with traditional methods, the method of the present invention can balance the gradients of different tasks. During the training process, the weight of the loss function is dynamically adjusted according to the gradient magnitude of different particle reaction equations. This avoids the situation in traditional methods where some tasks dominate the training and weak tasks are ignored, thereby improving the model's learning ability for multi-task processes and significantly improving the computational accuracy. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0071] Figure 2 This is a schematic diagram of the training of the chemical dynamics of environmentally friendly gases using the neural network of the present invention.

[0072] Figure 3 This is a schematic diagram of the adaptive loss function weight adjustment process in this invention;

[0073] Figure 4 This is a comparison chart of the main particle solution results of the method of the present invention in solving the chemical kinetics of the decomposition of pure C4F7N gas with those of conventional neural network differential equations;

[0074] Among them, (a) is a comparison diagram of C4F7N, (b) is a comparison diagram of CF, (c) is a comparison diagram of CF3CFCF2CN, (d) is a comparison diagram of CF3CFCN, (e) is a comparison diagram of CF2CN, and (f) is a comparison diagram of CF3CFCNCF2. Detailed Implementation

[0075] This invention proposes a method for calculating the chemical kinetics of environmentally friendly gases based on the constant differential equation, comprising four steps:

[0076] Step 1: First, determine the reaction pathway and reactant particles of the environmentally friendly gas reaction system, establish a chemical reaction kinetic model of the environmentally friendly gas, and generate a dataset;

[0077] Step 2: Then, based on the chemical kinetic model in Step 1, construct the corresponding neural ordinary differential equation framework, construct the loss function, and reduce the neural ordinary differential equation model and the loss function.

[0078] Step 3: Next, train the reduced neural ordinary differential model from Step 2, use an adaptive loss function to adjust and update the weight coefficients, and calculate the total network loss function.

[0079] Step 4: Finally, the gradient of the network after updating the weights in Step 3 is calculated through backpropagation. The gradient descent method is used to update the network parameters until the loss function decreases to a given threshold, thereby realizing the calculation of the chemical kinetic model of environmental gas and obtaining the result of the evolution of the concentration of each particle in the reaction system over time.

[0080] Step 1 is detailed below:

[0081] First, quantum chemistry and transition state theory are used to determine feasible reaction pathways for the environmentally friendly gas reaction system. The reaction rate constant for each reaction at a given temperature is calculated, and the reaction particles are identified. Based on the above, the following chemical kinetic model is constructed to describe M types of microscopic particles and K types of reactions:

[0082]

[0083] in, Let i be the chemical symbol for the i-th particle. and Let be the stoichiometric coefficients of the reactants and products of the i-th type of particle in the k-th reaction;

[0084] Then, based on the above reaction model, a set of ordinary differential equations can be constructed to describe the evolution of the particles participating in the reaction over time under a given initial reaction condition:

[0085]

[0086] in, Let be the concentration of the i-th type of particle. Let i be the stoichiometric coefficient of the i-th particle in the k-th reaction. Let k be the reaction rate of the k-th reaction;

[0087] Based on the established environmental gas chemical kinetic model, the initial and final values ​​of reactant particle concentrations over a given time period were obtained using a numerical method combining the Arrhenius equation. This constitutes the training and testing datasets.

[0088] Step 2 is detailed below:

[0089] First, based on the established environmental gas chemical kinetic model, construct the following framework of neural ordinary differential equations:

[0090]

[0091] In the formula, Y is the particle concentration vector. These are network parameters; a neural network (NN) is composed of given parameters. The defined neural network function; the network takes time t and particle concentration vector Y as input, learns the chemical kinetic source term function on the right-hand side, and then solves the system of ordinary differential equations to obtain the network output, that is, the change of particle concentration with time;

[0092] Then, construct the network's loss function based on the training task:

[0093]

[0094] In the formula, N is the number of data pairs, loss is an optional error calculation function, and Y* is the predicted value of the particle concentration vector. It is the starting time of the i-th time segment. It is the end time of the i-th time segment. Arrehenius is the source term of the chemical kinetic model obtained based on the Arrhenius formula, and NN in the formula represents the chemical kinetic source term function obtained based on the neural network.

[0095] Finally, the model of the ordinary differential equations of God is reduced:

[0096]

[0097] In the formula, It is the particle concentration reduced scale vector formed by subtracting the minimum value vector from the maximum value vector of all particle concentrations in the dataset. It is the time-reduced scale vector obtained by subtracting the minimum value from the maximum value of all times in the dataset;

[0098] The reduced loss function is calculated as follows:

[0099]

[0100] Step 3 is detailed below:

[0101] Step 3.1: Train the network based on the reduced neural ordinary differential model. First, initialize the task weight parameters corresponding to different reaction particles. Calculate the initial loss function ;

[0102] Step 3.2: In the t-th iteration of the network, calculate the loss function corresponding to different reactive particles. And calculate the loss function for each particle concentration relative to the neural network model parameters. gradient:

[0103]

[0104] Then calculate its L2 norm:

[0105]

[0106] Define gradient norm As an average baseline gradient norm that all tasks hope to achieve:

[0107]

[0108] Calculate the convergence rate of the i-th particle. :

[0109]

[0110] Based on this, the relative convergence velocity of the i-th particle is calculated. :

[0111]

[0112] This yields the target norm of the gradient norm for each particle's learning task. :

[0113]

[0114] In the formula It is the convergence rate adjustment factor, and its value range is... ;

[0115] Step 3.3, then based on the above calculations, we can obtain the weights. The loss function is shown in the following equation:

[0116]

[0117] Differentiating the loss function yields The corresponding gradient, thus affecting Update;

[0118] After the update The following normalization process is performed:

[0119]

[0120] Step 3.4: Finally, the total loss function of the network is calculated by weighting and summing the loss functions of different particles according to their weight coefficients, as shown in the following formula:

[0121]

[0122] In step 3, the updating of the loss function weight coefficients is decoupled from the updating of the neural network parameters in parallel. Depending on the actual situation, the loss function weights can be updated multiple times during a single neural network parameter iteration or after multiple neural network parameter iterations.

[0123] Step 4 is detailed below:

[0124] Calculate the total loss function relative to the network parameters The gradient is first determined by assessing the loss L at each time step for the hidden state. The gradient dependency, i.e., the associated state Its dynamic changes are described by the following ordinary differential equation:

[0125]

[0126] The calculation was performed by calling the ordinary differential equation solver again. In actual calculations, the final value can be simply obtained. Beginning, and Accompaniment Recalculate together Forward temporal evolution;

[0127] Next, calculate the parameters relative to the network. The gradient needs to be calculated using the following formula, which depends on... and Integrals:

[0128]

[0129] All used for solving , as well as The integrals are synthesized into an extended system of ordinary differential equations, which... Treating both and t as variables in the system of ordinary differential equations, we obtain the following equation:

[0130]

[0131] On this basis and Combining these, we obtain augmented variables with response ordinary differential equations and adjoint states:

[0132]

[0133]

[0134]

[0135]

[0136] Finally, solving this augmented equation yields the loss function L with respect to the network parameters. The gradient is used to update the network parameters using gradient descent until the loss function decreases to a given threshold, thus obtaining the result of the evolution of the concentration of each particle in the reaction system over time.

[0137] Please see Figure 1 As shown, the chemical kinetics of the decomposition of pure C4F7N gas at 3000K is taken as the research object. The simulation results of the concentration evolution of the main reacting particles over time are calculated using the method proposed in this invention, including the following steps:

[0138] Step 1: Establish a kinetic model for the decomposition chemical reaction of pure C4F7N gas, and generate training and testing datasets, as detailed below:

[0139] Step 1.1: First, based on quantum chemistry and transition state theory, it is determined that the pure C4F7N gas decomposition reaction system contains 50 main reaction pathways and 33 main reaction particles. At a given temperature T=3000K, the rate constants of each reaction are calculated using the Arrhenius equation. Based on the obtained data of the pure C4F7N gas decomposition reaction system, the following chemical kinetic model is constructed:

[0140]

[0141] in, Let i be the chemical symbol for the i-th particle. and Let be the stoichiometric coefficients of the reactants and products of the i-th type of particle in the k-th reaction;

[0142] Then, based on the above reaction model, a set of ordinary differential equations can be constructed to describe the evolution of the particles participating in the reaction over time under a given initial reaction condition:

[0143]

[0144] in, Let be the concentration of the i-th type of particle. Let i be the stoichiometric coefficient of the i-th particle in the k-th reaction. Let k be the reaction rate of the k-th reaction;

[0145] Step 1.2: Select the time interval as [10 -13 10 -2 [seconds], by combining the traditional numerical method with the Arrhenius formula, the initial and final values ​​of the reaction particle concentration were calculated. , generating 500 pairs of training data and 300 pairs of test data.

[0146] Step 2: Construct the corresponding neural ordinary differential equation framework based on the chemical kinetic model in Step 1, such as... Figure 2As shown, a loss function is constructed, and the ordinary differential equation model and loss function are reduced, as detailed below:

[0147] Step 2.1: Construct an ordinary differential equation of the following form based on the chemical kinetic model:

[0148]

[0149] In the formula, Y is the particle concentration vector. These are network parameters; a neural network (NN) is composed of given parameters. Defined neural network function;

[0150] Step 2.2: Construct a structure with time t and particle concentration vector Y as input. The output neural network has 3 hidden layers, each with 50 neurons, and uses the Adam optimizer with a learning rate of 10. -2 The number of training sessions is set to 50,000.

[0151] Step 2.3: Construct the network's loss function based on the training task as follows:

[0152]

[0153] In the formula, N is the number of data pairs, and Y* is the predicted value of the particle concentration vector. It is the starting time of the i-th time segment. It is the end time of the i-th time segment. Arrehenius is the source term of the chemical kinetic model obtained based on the Arrhenius formula, NN is the source term function of the chemical kinetic model obtained based on the neural network, and MAE error is used as the error calculation function loss.

[0154] Step 2.4: Reduce the ordinary differential equation model:

[0155]

[0156] In the formula, It is the particle concentration reduced scale vector formed by subtracting the minimum value vector from the maximum value vector of all particle concentrations in the dataset. It is the time-reduced scale vector obtained by subtracting the minimum value from the maximum value of all times in the dataset;

[0157] Step 2.5, Reduction processing loss function calculation formula:

[0158]

[0159] Step 3: Train the reduced neural network frequent differential model from Step 2, using an adaptive loss function to adjust and update the weight coefficients, and calculate the total network loss function, such as... Figure 3 As shown, details are as follows:

[0160] Step 3.1: Initialize the task weight parameters corresponding to each reaction particle. Calculate the initial loss function ;

[0161] Step 3.2: Set the task weight update frequency to once every 10 network iterations. At the t-th iteration of the network, calculate the loss function corresponding to different reaction particles. And calculate the loss function for each particle concentration relative to the neural network model parameters. gradient:

[0162]

[0163] And calculate its L2 norm:

[0164]

[0165] Step 3.3: Define the gradient norm As an average baseline gradient norm that all tasks hope to achieve:

[0166]

[0167] Step 3.4: Calculate the convergence rate of the i-th particle. :

[0168]

[0169] Based on this, the relative convergence velocity of the i-th particle is calculated. :

[0170]

[0171] Step 3.5: Set the convergence rate adjustment factor The target norm is 1.5, used to obtain the gradient norm for each particle learning task. :

[0172]

[0173] Step 3.6: Based on the above calculations, the weights can be obtained. Loss function:

[0174]

[0175] Step 3.7: Take the derivative of the loss function to obtain... The corresponding gradient, thus affecting Update, and after the update... The following normalization process is performed:

[0176]

[0177] Step 3.8: The total loss function of the network is calculated by weighting and summing the loss functions of different particles according to the weight coefficients, as shown in the following formula:

[0178]

[0179] Step 4: Finally, calculate the gradient of the network through backpropagation, and update the network parameters using gradient descent until the loss function decreases to a given threshold, as detailed below:

[0180] Step 4.1: Determine the loss function for the hidden state at each time step. Dynamic changes in gradient dependencies:

[0181]

[0182] The calculation of the following formula depends on and Integrals:

[0183]

[0184] Step 4.2: Use all the solutions... , as well as The integrals are synthesized into an extended system of ordinary differential equations, which... Treating both and t as variables in the system of ordinary differential equations, we obtain the following equation:

[0185]

[0186] On this basis and Combining these, we obtain augmented variables with response ordinary differential equations and adjoint states:

[0187]

[0188]

[0189]

[0190]

[0191] Step 4.3: Solve the augmented equation to obtain the loss function L with respect to the neural network parameters. The gradient is used to update the network parameters using gradient descent.

[0192] Step 4.4: Repeat steps 4.1-4.3 and observe the loss function decrease to a given threshold;

[0193] Step 4.5: Obtain the output of the neural network, which is the simulation result of the evolution of the concentration of each particle in the corresponding chemical kinetic reaction system over time.

[0194] The comparison figures for the training results of the improved neural ordinary differential equations proposed in this invention and the training results of the conventional neural ordinary differential equations are shown below. Figure 4 As shown, due to the large number of particles involved in the reaction, only some of the main particle training results are displayed. Figure 4 It can be seen that although conventional neural ordinary differential equations can overcome certain rigidity problems, capture the concentration distribution evolution of most particles, and provide relatively accurate predictions, certain errors still occur in the prediction results during the initial stage of the reaction and during periods of drastic particle changes. The improved neural ordinary differential equation method proposed in this invention can provide accurate predictions under all particle concentration evolution conditions, thus improving the computational accuracy of rigid chemical kinetic models.

[0195] Next, the number of iterations required for the conventional and the improved neural frequent differential equations proposed in this invention to reach a given loss and relative L2 error threshold during training were compared. The results are summarized in Table 1. The data in the table shows that the improved neural frequent differential equations achieved superior training performance compared to the conventional neural frequent differential equations, which still failed to converge to 10 even after reaching the set maximum number of iterations. -3 The improved neural ordinary differential equation (NDE) model converges to values ​​below the specified loss threshold and L2 error threshold. Furthermore, it was found that the improved NDE model significantly outperforms the conventional NDE model in training convergence speed. For a set loss threshold of 10... -1 10 -2 The improved neural ordinary differential equations converged at rates 1.4 and 1.5 times faster than the conventional neural ordinary differential equations, respectively; similar results were observed with the L2 error, for a set error threshold of 10. -1 10 -2 The improved neural ordinary differential equations (NDEs) show convergence rates that are 1.4 times and 1.8 times higher than those of the conventional NDEs, respectively. These results demonstrate that the improved NDE method presented in this example can accelerate the training of environmental gas chemical kinetic models and improve computational efficiency.

[0196]

[0197] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0198] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0199] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0200] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0201] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for calculating the chemical kinetics of environmentally friendly gases based on the constant differential equation, characterized in that... Includes the following steps: Step 1: Based on the feasible reaction pathways and reaction particles of the environmentally friendly gas reaction system, construct a chemical reaction kinetic model of the environmentally friendly gas, and construct a set of ordinary differential equations of chemical kinetics based on the environmentally friendly gas chemical reaction kinetic model; Step 2: Based on the chemical kinetic ordinary differential equations, construct the neural ordinary differential equations and loss function, and reduce the neural ordinary differential equations and loss function to obtain the neural ordinary differential model. Step 3: Train the neural network based on the training and testing datasets, and use an adaptive loss function to adjust and update the weight coefficients, and calculate the total network loss function; Step 4: Calculate the gradient of the updated weight coefficients network through backpropagation, and update the network parameters using the gradient descent method until the loss function decreases to a given threshold, thus obtaining the result of the evolution of the concentration of each particle in the reaction system over time. In step 2, based on the set of chemical kinetic ordinary differential equations, the neural network's ordinary differential equations are constructed as follows: ， Where Y is the particle concentration vector, and t is time. These are network parameters; a neural network (NN) consists of parameters. Defined neural network function; Construct the loss function of the neural network based on the training task. ， Where N is the number of data pairs, loss is the error calculation function, and Y* is the predicted particle concentration vector value. It is the starting time of the i-th time segment. It is the end time of the i-th time segment. Arrehenius is the source term of the chemical kinetic model obtained based on the Arrhenius formula, and NN is the source term function of the chemical kinetic model obtained based on the neural network. By reducing the constant differential equation of God, we obtain: ， Among them, Y scale Y is the vector of maximum particle concentrations across all particles in the dataset. max Subtract the minimum value vector Y min The particle concentration reduced scale vector, t scale It is the time-reduced scale vector obtained by subtracting the minimum value t0 from the maximum value t1 of all times in the dataset; After reducing the loss function, we get: ； Step 3 specifically includes the following steps. Step 31: Train the God frequent differential model and initialize the task weight parameters corresponding to different reaction particles. Calculate the initial loss function ; Step 32: At the t-th iteration of the network, calculate the loss function corresponding to different reaction particles. And calculate the loss function for each particle concentration relative to the network parameters. gradient, ， Then calculate its L2 norm. ， Define gradient norm As the average baseline gradient norm that all tasks hope to achieve ， Calculate the convergence rate of the i-th particle. , ， Then calculate the relative convergence velocity of the i-th particle. , ， The target norm for obtaining the gradient norm of each particle learning task , ， in, It is the convergence rate adjustment factor, and its value range is... ; Step 33, obtain information about the weights. loss function, ， Differentiating the loss function yields The corresponding gradient, thus affecting Update; Step 34: The total loss function of the network is calculated by weighting and summing the loss functions of different particles according to the weight coefficients.

2. The method as described in claim 1, characterized in that: The specific process of step 1 is as follows: The feasible reaction pathways of the environmentally friendly gas reaction system are determined using quantum chemistry and transition state theory. The reaction rate constants for each reaction at a given temperature are calculated, and the reaction particles are identified. The following chemical kinetic model is constructed to describe a chemical reaction involving M types of reacting particles and K types of reactions. ， in, Let i be the chemical symbol for the i-th particle. and These are the stoichiometric coefficients of the reactants and products of the i-th particle in the k-th reaction, respectively. A set of ordinary differential equations is constructed to describe the evolution of the particles participating in the reaction over time under a given initial reaction condition. ， in, Let be the concentration of the i-th type of particle. Let i be the stoichiometric coefficient of the i-th particle in the k-th reaction. Let be the reaction rate of the k-th reaction.

3. The method as described in claim 1, characterized in that: In step 3, the initial and final values ​​of the concentration of reactive particles over a given time period are obtained using the numerical method of the Arrhenius formula, forming training and testing datasets.

4. The method as described in claim 1, characterized in that: In step 3, the updating of the loss function weight coefficients is decoupled from the updating of network parameters in parallel. The loss function weights are updated multiple times during a single network parameter iteration update or after multiple network parameter iteration updates.

5. The method as described in claim 1, characterized in that: Step 4 specifically includes the following steps. Step 41, determine the loss L for the hidden state at each time step. The gradient dependency, i.e., the associated state Its dynamic changes are described by the following ordinary differential equation: ， Calculate relative to network parameters gradient, ， Step 42, use all the solutions , as well as The integrals are synthesized into an extended system of ordinary differential equations, which... Treating both and t as variables in the system of ordinary differential equations, we obtain the following equation: ， On this basis and By combining these, we obtain augmented variables that have a system of ordinary differential equations for the response and the adjoint state. ， ， ， ， Step 43: Solve this augmented equation to obtain the loss function L with respect to the network parameters. The gradient is used to update the network parameters using gradient descent. Step 44: Repeat steps 41-43 until the loss function drops to a given threshold; Step 45 yields the network output, which is the result of the evolution of the concentration of each particle in the reaction system over time.

Images