A combustion mechanism optimization method based on a generative neural network

Abstract

The application discloses a combustion mechanism optimization method based on a generative neural network, and comprises the following steps: step S1, optimizing mechanism dynamics parameters according to a method; step S2, analyzing sensitivity, constructing a DNN and optimizing a DRM19 mechanism according to the optimized mechanism dynamics parameters; and step S3, drawing a conclusion according to the results of analyzing sensitivity, constructing the DNN and optimizing the DRM19 mechanism. The application solves the problem that in the mechanism simplification process, with the reduction of the number of species and reactions, the obtained simplified mechanism will cause certain loss in the prediction accuracy of IDT and LFS combustion characteristics.

Description

Technical Field

[0001] This invention belongs to the field of combustion mechanism optimization methods, specifically relating to a combustion mechanism optimization method based on a generative neural network. Background Technology

[0002] Detailed mechanistic models of fuels typically involve numerous species and reactions, making them costly in computational fluid dynamics (CFD) simulations. To facilitate the application of fuel mechanisms in numerical simulations of combustion chambers, various mechanism simplification methods have been developed in recent years, including lumped and skeleton simplification methods. Lumped methods reduce the size of mechanisms by combining similar species and reaction pathways. As a primary simplification method, skeleton simplification addresses complex couplings between species or reactions by eliminating unnecessary species and reactions, primarily including directed relation graph (DRG) methods and a series of subsequently proposed derivative methods.

[0003] Methane was chosen as the focus of this study because it is a key intermediate in the oxidation and pyrolysis of large hydrocarbons and is widely used as a fuel in practical engine applications. Due to its importance in these contexts, many researchers have simplified and experimentally verified the methane combustion mechanism. Kazakov et al. simplified GRI Mech 1.2 using ignition delay time (IDT) and laminar flame velocity (LFS) as indicators to obtain a simplified model of DRM22, which includes 24 substances and 108 reactions. Based on the results of sensitivity analysis, after removing species with near-zero sensitivity, a simplified mechanism of DRM19, including 21 species and 84 reactions, was obtained. Lu et al., using IDT as the standard, derived a skeletal mechanism consisting of 39 substances and 231 reactions from the methane Mech_56.54 mechanism (113 substances and 710 reactions) based on the Directed Relationship Graph of Error Propagation (DRGEP) method, with a maximum relative error of 5.88% compared to the detailed mechanism. Gou et al. simplified the methane mechanism using an improved path flux analysis and multi-generational analysis (IMPFA) method, reducing the number of 24 species involved, and verified its reliability through changes in IDT and species concentration under air conditions. Regarding the combustion characteristics of methane, Lifshitz et al. measured the IDT of argon-diluted methane within the range of equivalence ratio Φ∈[0.5, 2.0], T∈[1500 K, 2150 K], and p∈[2 atm, 10 atm]. Hu et al. measured the IDT of methane under ambient air conditions at equivalence ratios of 1300–1950 K, 1–10 atm, and 0.5–2.0. Furthermore, they measured LFS within the range of p = 0.1–0.5 MPa, T = 300–443 K, and Φ = 0.6–1.4.

[0004] During the mechanistic simplification process, the reduction in the number of species and reactions leads to a certain loss in the predictive accuracy of combustion characteristics such as IDT and LFS. Therefore, it is necessary to optimize the simplified model to enhance its consistency with experimental data, thereby ensuring the accuracy of CFD simulations. The kinetic model optimization process can be viewed as solving an inverse problem with the help of sensitivity analysis, uncertainty analysis, and data regression. In recent years, there has been much work on optimizing kinetic parameters. Elliott et al. extensively studied the applicability of genetic algorithms (GA) in optimizing detailed kinetic models. Ryu et al. selected a certain number of important reactions based on temperature sensitivity analysis using a genetic algorithm and optimized the kinetic model of F-24 jet fuel using IDT as the index. In addition, Ji et al. optimized the kinetic parameters of simplified combustion mechanisms of methane and n-heptane using stochastic gradient descent (SGD) with IDT as the optimization objective. The optimized models showed high consistency with the simulation results of detailed combustion mechanisms in terms of predictive performance. It is worth noting that almost all optimization work focuses on IDT as the sole objective, while IDT is only the most easily achievable of the common combustion characteristics.

[0005] This study explores a novel approach to optimizing combustion models using Generative Neural Networks (GNNs). Generative Neural Networks are models designed to generate new data samples and are commonly used in fields such as image and text generation. Recently, these networks have shown significant advantages in solving multivariate functions, particularly for complex, nonlinear, and high-dimensional problems. To our knowledge, however, it has not yet been applied to combustion mechanism optimization. We utilize GNNs to generate kinetic parameters and employ a custom loss function to guide it in generating better parameters. The advantage of this approach is that researchers can optimize kinetic models based solely on experimental data, even without extensive theoretical background in combustion kinetics. To demonstrate the optimization effectiveness of GNNs, this study selects the simplified methane model DRM19 as a case study. Compared to the IDT of methane under ambient air conditions, Figure 1 The simulation results using the DRM19 mechanism are shown. Figure 1 The symbols in the middle represent experimental data, and the solid lines represent simulated data. It can be observed that, under the condition of Φ=2.0, the predicted IDT deviates significantly from the experimental data, with a maximum relative error reaching 116.7%. Furthermore, as... Figure 2 As shown, the DRM19 mechanism overestimates the LFS at low stoichiometric ratios under all operating conditions, with a maximum relative error of 51.67%. To minimize the prediction error of the DRM19 mechanism, this study targets the IDT and LFS in air and uses the GNN method to optimize and simplify the dynamic parameters of the mechanism. Figure 2 The symbols in the figure represent experimental data, and the solid lines represent simulated data. Referring to previous work on kinetic parameter optimization, this work only optimizes the pre-exponential factors of the key reactions to reduce the dimensionality of the optimization problem. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a combustion mechanism optimization method based on generative neural networks. This method solves the problem that, during the mechanism simplification process, as the number of species and reactions decreases, the obtained simplified mechanism suffers a certain loss in the prediction accuracy of IDT and LFS combustion characteristics.

[0007] The technical solution provided by this invention is as follows:

[0008] A combustion mechanism optimization method based on generative neural networks includes the following steps:

[0009] Step S1: Optimize the mechanistic dynamic parameters according to the method;

[0010] Step S2: Based on the optimized mechanism dynamic parameters, analyze the sensitivity, construct the DNN, and optimize the DRM19 mechanism;

[0011] Step S3: Based on the results of sensitivity analysis, DNN construction, and DRM19 mechanism optimization, draw conclusions.

[0012] Preferably, step S1 includes the following steps:

[0013] Step S11: Perform sensitivity analysis;

[0014] Step S12: Based on the sensitivity analysis results, establish a deep neural network;

[0015] Step S13: Generative network based on deep neural network.

[0016] Preferably, the conclusion of step S3 is as follows:

[0017] Using the PyTorch framework, corresponding program code was developed to implement the optimization process;

[0018] The simplified mechanism of methane combustion, DRM19, was selected, and 17 key reactions affecting IDT and LFS were identified. The corresponding pre-exponential factors were optimized. The optimized model DRM19-opt reduced the MAE of IDT from 32.80% to 14.41% and the MAE of LFS from 17.33% to 5.33%.

[0019] The simulation results of the optimized model are in good agreement with the experimental values ​​independent of the training data, proving the effectiveness of GNN.

[0020] Preferably, step S11 specifically involves: in order to identify the key reactions affecting IDT during combustion, a local sensitivity analysis of the mechanism was performed under experimental temperature, pressure, and equivalence ratio conditions, and the calculation formula is as follows:

[0021]

[0022] Where τ represents IDT, Representing the The reaction rate of one reaction, 2 The reaction rate is doubled; a negative value of S indicates that the corresponding reaction promotes the ignition process.

[0023] The sensitivity coefficient of LFS is calculated using the normalized sensitivity method in Cantera. The formula is as follows:

[0024]

[0025] in, Represents LFS, Representative When there are minor changes The changes.

[0026] Preferably, step S12 includes the following steps:

[0027] Step S121: Define the input vector of the DNN as x = {Φ, p, T, Ai}, where Ai represents the exponential factor of the i-th reaction to be optimized;

[0028] Step S122: Standardize the dataset input using Z-scores to make the data dimensionless and to unify the magnitudes of different variables. Specifically, the Z-score is:

[0029]

[0030] Where, x, x mean x std These represent the original data, mean, and standard deviation, respectively. In this work, the output target value y is determined by IDT and LFS.

[0031] Step S123: Train the dataset using a neural network with L layers, where the calibrated linear unit is used as the activation function and the mean absolute error is used as the loss function. The formula for calculating the loss function is:

[0032]

[0033] in, This represents the predicted value of the DNN. This represents the corresponding actual value.

[0034] Preferably, step S13 specifically includes:

[0035] By coupling with a pre-trained DNN, combustion characteristics IDT and LFS corresponding to kinetic parameters are quickly obtained, and a custom GNN loss function is used to guide the generation of better parameters.

[0036] The loss function of GNN is expressed as:

[0037]

[0038] in, and These are hyperparameters representing the importance of IDT and LFS. and These are simulated values ​​obtained through a DNN. and These are experimental values.

[0039] Preferably, the analytical sensitivity of step S2 includes the following steps:

[0040] Step S211: Local sensitivity analysis was performed under the conditions of Φ = 2.0, T = 1600 K, p = 1, 5, and 10 atm;

[0041] Step S212: Sensitivity analysis of LFS was performed at p = 0.1 MPa, Φ = 0.7, T = 373 and 443 K.

[0042] Preferably, step S2, constructing the DNN, includes the following steps:

[0043] Step S221: According to the experimental working conditions of IDT, the experimental conditions are: Φ∈[0.5,2.0], p∈[1 atm,10 atm], T∈[1300 K,2000 K], sampling is performed within the range of variation of the pre-exponential factors of 17 key reactions;

[0044] Step S222: Simulate all inputs using Cantera to obtain the corresponding IDT and LFS;

[0045] Step S223: The input data and simulation results constitute a dataset, which is divided into a training set, a validation set, and a test set in a ratio of 0.7:0.2:0.1.

[0046] Preferably, the optimized DRM19 mechanism in step S2 includes the following steps:

[0047] Step S231: Based on the trained DNN, GNN is used to optimize the parameters of 17 key reactions in the DRM19 dynamic model;

[0048] Step S232: During training, the GNN receives a set of random 17-dimensional noise and generates a set of 17-dimensional data as potential optimization parameters.

[0049] Step S233: After combining these generated parameters with experimental conditions, use a pre-trained DNN to quickly predict the IDT and LFS of the dynamic model corresponding to these parameters;

[0050] Step S234: Calculate the loss using a custom GNN loss function and provide feedback to the GNN to guide it in generating better parameters;

[0051] Step S235: Place λ IDT and λ LFS The value is adjusted within the range of 0.2 to 0.8;

[0052] Step S236: After multiple tests, the results show that when λ IDT = 0.7 and λ LFS The optimized dynamic model performs best when the value is 0.3.

[0053] The beneficial effects of the combustion mechanism optimization method based on generative neural networks in this invention are as follows:

[0054] 1. The method of the present invention is both simple and efficient, requiring only the establishment of the relationship between kinetic parameters and combustion characteristics through DNN.

[0055] 2. This invention proposes a novel mechanism optimization method by using a simple generative network for optimization. Attached Figure Description

[0056] To more clearly illustrate the purpose, design concept, and innovation of the combustion mechanism optimization method based on generative neural networks proposed in this invention, the invention will be described in detail below with reference to the accompanying drawings and tables.

[0057] Figure 1 The figures show experimental and simulation data of the methane-air mixture IDT of the present invention.

[0058] Figure 2 The figures show experimental and simulated data of the methane-air mixture LFS of the present invention.

[0059] Figure 3 This is a schematic diagram of the GNN optimization of the present invention.

[0060] Figure 4 This is a schematic diagram of the DNN network structure of the present invention.

[0061] Figure 5 This is a sensitivity analysis diagram of the ignition delay time of the present invention.

[0062] Figure 6 This is a sensitivity analysis diagram of the laminar flame velocity of the present invention.

[0063] Figure 7 The DNN of this invention IDT The training process diagram.

[0064] Figure 8(a) shows the DNN of the present invention. IDT A comparison chart of predicted values ​​and Cantera calculated values ​​at φ=0.5.

[0065] Figure 8(b) shows the DNN of the present invention. IDT A comparison chart of predicted values ​​and Cantera calculated values ​​at φ=1.0.

[0066] Figure 8(c) shows the DNN of the present invention. IDT A comparison chart of predicted values ​​and Cantera calculated values ​​at φ=2.0.

[0067] Figure 9(a) shows the DNN of the present invention. LFS A comparison plot of predicted values ​​and Cantera calculated values ​​at p=0.1MPa.

[0068] Figure 9(b) shows the DNN of the present invention. LFS A comparison chart of predicted values ​​and Cantera calculated values ​​at p=0.2 MPa.

[0069] Figure 10(a) is a comparison of the IDT simulation results between the optimized DRM19-opt model and DRM19 of the present invention with φ=0.5.

[0070] Figure 10(b) is a comparison of the IDT simulation results between the optimized DRM19-opt model and DRM19 of the present invention with φ=1.0.

[0071] Figure 10(c) is a comparison of the IDT simulation results between the optimized DRM19-opt model and DRM19 of the present invention (φ=2.0).

[0072] Figure 11(a) is a comparison of the LFS simulation results between the optimized DRM19-opt model and DRM19 according to the present invention (p=0.1MPa).

[0073] Figure 11(b) is a comparison of the LFS simulation results between the optimized DRM19-opt model and DRM19 according to the present invention (p=0.2MPa).

[0074] Figure 11(c) is a comparison of the LFS simulation results between the optimized DRM19-opt model and DRM19 according to the present invention (p=0.1MPa, T=300K).

[0075] Figure 11(d) is a comparison of the LFS simulation results between the optimized DRM19-opt model and DRM19 according to the present invention, p=0.5MPa, T=443K. Detailed Implementation

[0076] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0077] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0078] A combustion mechanism optimization method based on generative neural networks includes the following steps:

[0079] Step S1: Optimize the mechanistic dynamic parameters according to the method;

[0080] Step S2: Based on the optimized mechanism dynamic parameters, analyze the sensitivity, construct the DNN, and optimize the DRM19 mechanism;

[0081] Step S3: Based on the results of sensitivity analysis, DNN construction, and DRM19 mechanism optimization, draw conclusions.

[0082] Step S1 of this implementation plan includes the following steps:

[0083] Step S11: Perform sensitivity analysis;

[0084] Step S12: Based on the sensitivity analysis results, establish a deep neural network;

[0085] Step S13: Generative network based on deep neural network.

[0086] The specific conclusion of step S3 in this implementation plan is as follows:

[0087] Using the PyTorch framework, corresponding program code was developed to implement the optimization process;

[0088] The simplified mechanism of methane combustion, DRM19, was selected, and 17 key reactions affecting IDT and LFS were identified. The corresponding pre-exponential factors were optimized. The optimized model DRM19-opt reduced the MAE of IDT from 32.80% to 14.41% and the MAE of LFS from 17.33% to 5.33%.

[0089] The simulation results of the optimized model are in good agreement with the experimental values ​​independent of the training data, proving the effectiveness of GNN.

[0090] Step S11 of this implementation plan specifically involves: In order to identify the key reactions affecting IDT during combustion, a local sensitivity analysis of the mechanism was performed under experimental temperature, pressure, and equivalence ratio conditions. The calculation formula is as follows:

[0091]

[0092] Where τ represents IDT, Representing the The reaction rate of one reaction, 2 The reaction rate is doubled; a negative value of S indicates that the corresponding reaction promotes the ignition process.

[0093] The sensitivity coefficient of LFS is calculated using the normalized sensitivity method in Cantera. The formula is as follows:

[0094]

[0095] in, Represents LFS, Representative When there are minor changes The changes.

[0096] Step S12 of this implementation plan includes the following steps:

[0097] Step S121: Define the input vector of the DNN as x = {Φ, p, T, Ai}, where Ai represents the exponential factor of the i-th reaction to be optimized;

[0098] Step S122: Standardize the dataset input using Z-scores to make the data dimensionless and to unify the magnitudes of different variables. Specifically, the Z-score is:

[0099]

[0100] Where, x, x mean x std These represent the original data, mean, and standard deviation, respectively. In this work, the output target value y is determined by IDT and LFS.

[0101] Step S123: Train the dataset using a neural network with L layers, where the calibrated linear unit is used as the activation function and the mean absolute error is used as the loss function. The formula for calculating the loss function is:

[0102]

[0103] in, This represents the predicted value of the DNN. This represents the corresponding actual value.

[0104] Step S13 of this implementation plan is as follows:

[0105] By coupling with a pre-trained DNN, combustion characteristics IDT and LFS corresponding to kinetic parameters are quickly obtained, and a custom GNN loss function is used to guide the generation of better parameters.

[0106] The loss function of GNN is expressed as:

[0107]

[0108] in, and These are hyperparameters representing the importance of IDT and LFS. and These are simulated values ​​obtained through a DNN. and These are experimental values.

[0109] The sensitivity analysis in step S2 of this implementation plan includes the following steps:

[0110] Step S211: Local sensitivity analysis was performed under the conditions of Φ = 2.0, T = 1600 K, p = 1, 5, and 10 atm;

[0111] Step S212: Sensitivity analysis of LFS was performed at p = 0.1 MPa, Φ = 0.7, T = 373 and 443 K.

[0112] The construction of the DNN in step S2 of this implementation scheme includes the following steps:

[0113] Step S221: According to the experimental working conditions of IDT, the experimental conditions are: Φ∈[0.5,2.0], p∈[1 atm,10 atm], T∈[1300 K,2000 K], sampling is performed within the range of variation of the pre-exponential factors of 17 key reactions;

[0114] Step S222: Simulate all inputs using Cantera to obtain the corresponding IDT and LFS;

[0115] Step S223: The input data and simulation results constitute a dataset, which is divided into a training set, a validation set, and a test set in a ratio of 0.7:0.2:0.1.

[0116] The optimized DRM19 mechanism in step S2 of this implementation scheme includes the following steps:

[0117] Step S231: Based on the trained DNN, GNN is used to optimize the parameters of 17 key reactions in the DRM19 dynamic model;

[0118] Step S232: During training, the GNN receives a set of random 17-dimensional noise and generates a set of 17-dimensional data as potential optimization parameters.

[0119] Step S233: After combining these generated parameters with experimental conditions, use a pre-trained DNN to quickly predict the IDT and LFS of the dynamic model corresponding to these parameters;

[0120] Step S234: Calculate the loss using a custom GNN loss function and provide feedback to the GNN to guide it in generating better parameters;

[0121] Step S235: Place λ IDT and λ LFS The value is adjusted within the range of 0.2 to 0.8;

[0122] Step S236: After multiple tests, the results show that when λ IDT = 0.7 and λ LFS The optimized dynamic model performs best when the value is 0.3.

[0123] When this implementation plan is implemented,

[0124] 1. Method

[0125] The corresponding process for optimizing the mechanistic dynamic parameters is as follows: Figure 3 As shown. Based on the simulation results, local sensitivity analysis was performed on conditions with larger errors to determine the response that needs optimization. Sampling was then performed. The IDT and LFS of the kinetic model were calculated using Cantera software. A DNN was trained using a dataset containing sampled data and corresponding targets, with the aim of enabling the network to accurately predict the targets. During optimization, the GNN created optimization parameters from a set of noise and combined them with the experimental conditions as input to the DNN, which then provided the predicted targets under the corresponding conditions. Subsequently, a custom loss function guided the GNN to generate better parameters.

[0126] This invention uses the PyTorch framework to build DNN and GNN, and develops corresponding program code to implement this optimization process.

[0127] 1.1 Sensitivity Analysis

[0128] To identify the key reactions affecting IDT during combustion, a local sensitivity analysis of the mechanism was performed under experimental temperature, pressure, and equivalence ratio conditions. The sensitivity coefficients were calculated using the method proposed by Kumar et al., as shown in the following formula:

[0129] (1)

[0130] Where τ represents IDT, k i 2k represents the reaction rate of the i-th reaction. i The reaction rate was doubled. A negative value for S indicates that the corresponding reaction promoted the ignition process, and vice versa. The normalized sensitivity method in Cantera was used to calculate the sensitivity coefficient of LFS.

[0131] (2)

[0132] in, Representing LFS, Representative When there are minor changes The changes.

[0133] 1.2 Deep Neural Networks

[0134] We define the input vector of the DNN as x = {Φ, p, T, Ai}, where Ai represents the exponential factor of the i-th reaction to be optimized. The dataset input is standardized using the Z-score (Equation 3) to make the data dimensionless and to unify the magnitudes of the different variables.

[0135]

[0136] Where x, x mean , x std These represent the original data, mean, and standard deviation, respectively. In this work, the output target value y is calculated using IDT and LFS. Figure 4 A schematic diagram of a DNN used to process IDT is given. Similar networks are also used to establish the relationship between dynamic parameters and LFS.

[0137] We use a neural network with L layers to train the dataset, where the Rectified Linear Unit (ReLU) is used as the activation function and the Mean Absolute Error (MAO) is used as the loss function (Equation 4).

[0138]

[0139] in, This represents the predicted value of the DNN. This represents the corresponding actual value.

[0140] 1.3 Generative Networks

[0141] GNN employs a general deep fully connected neural network architecture, taking a set of random numbers as input and optimized kinetic parameters as output. By coupling with a pre-trained DNN, it can quickly obtain combustion characteristics (IDT and LFS) corresponding to the kinetic parameters, and uses a custom GNN loss function to guide the generation of better parameters. The GNN loss function can be expressed as:

[0142] (5)

[0143] in, and Hyperparameters representing the importance of IDT and LFS. and These are simulated values ​​obtained through a DNN. and These are experimental values.

[0144] 2 Results and Discussion

[0145] 2.1 Sensitivity Analysis

[0146] Because the DRM19 model overestimates IDT at high equivalence ratios and high temperatures, local sensitivity analyses were performed under the conditions of Φ = 2.0, T = 1600 K, and p = 1, 5, and 10 atm. The corresponding results are as follows: Figure 5 As shown, under the three pressure conditions, CH3 + O2 = O + CH3O and H + O2 = O + OH are key reactions that promote ignition because they are branching reactions that consume O2. Conversely, the reactions 2CH3(+M) = C2H6(+M) and H + CH4 = CH3 + H2 both inhibit ignition. Furthermore, other reactions related to CH3 and CH3O species also promote or inhibit ignition.

[0147] Sensitivity analysis of LFS was performed at p = 0.1 MPa, Φ = 0.7, T = 373 and 443 K. The results are as follows: Figure 6 As shown, chain branching reactions that generate hydrogen, hydroxyl radicals, and other highly reactive free radical species promote LFS, such as...

[0148] H + O₂ = O + OH, CO + OH = CO₂ + H. In contrast, reactions that consume reactive free radical species have a negative impact on LFS, such as H + O₂ + H₂O = HO₂ + H₂O, CH₃ + H(+M) = CH₄(+M). Among the top 10 sensitive reactions affecting LFS and IDT, three reactions were repeated, and a total of 17 reactions were ultimately selected for optimization.

[0149] 2.2 Construction of DNN

[0150] Based on the experimental operating conditions of the IDT (Φ∈[0.5,2.0], p∈[1 atm, 10 atm], T∈[1300 K, 2000 K]), sampling was performed within the range of variation of the pre-exponential factors of 17 key responses. Cantera was used to simulate all inputs to obtain the corresponding IDT and LFS. The input data and simulation results constituted a dataset, which was divided into training, validation, and test sets in a ratio of 0.7:0.2:0.1.

[0151] Figure 7 DNN was displayed IDT The evolution of the loss function on the training and validation sets during training. After 298 training iterations, the DNN constructed in this work... IDT Stop when the mean absolute error (MAE) on the test set is less than 10⁻⁵. Furthermore, to validate the DNN... IDT To assess the general predictive ability of the DNN, we randomly generate a set of data within the sampling range and then apply it to the DNN. IDT The results are compared with the true values ​​obtained from Cantera simulations. It should be noted that these specific data points were not included in the training set. As shown in Figure 8, the DNN... IDT The predicted values ​​are highly consistent with the data simulated by Cantera. Dashed line: DNN IDT Predicted value, point: Cantera calculated value. This indicates that the DNN... IDT It exhibits reliable performance in predicting IDTs under different operating conditions and can be used as an alternative network for computing IDTs. The same procedure is used to train the DNN. LFS DNN LFS Figure 9 shows a comparison of the simulation results with Cantera. In Figure 9, the dashed line represents the DNN. LFS Predicted value, point: Cantera calculated value.

[0152] 2.3DRM19 Mechanism Optimization

[0153] Based on a pre-trained DNN, a GNN is used to optimize the parameters of 17 key responses in the DRM19 dynamics model. During training, the GNN receives a set of random 17-dimensional noise and then generates a set of 17-dimensional data as potential optimization parameters. After combining these generated parameters with experimental conditions, the pre-trained DNN is used to quickly predict the IDT and LFS of the dynamics model corresponding to these parameters. Then, a custom loss function (Equation 5) is used to calculate the loss and provide feedback to the GNN, guiding it to generate better parameters. We will use λ IDT and λ LFS The value was adjusted to be within the range of 0.2 to 0.8. After multiple tests, the final results showed that when λ... IDT = 0.7 and λ LFS The optimized dynamic model performs best when the value is 0.3.

[0154] After optimization, Figure 10 compares the IDT simulation results and experimental data of the optimized DRM19-opt model and the original DRM19. Solid line: DRM19-opt. Dashed line: DRM19. Dots: Experimental data. Under the experimental conditions, the consistency between the DRM19-opt model and the experimental data is significantly improved, with the average error decreasing from 32.80% to 14.41%. Especially under the condition of Φ = 2.0, the optimization effect of GNN is obvious, and the simulation results of DRM19-opt are significantly improved. Figure 11 shows the comparison between the simulation results and experimental values ​​of LFS for DRM19 opt and DRM19. Solid line: DRM19-opt. Dashed line: DRM19. Dots: Experimental data. It can be seen that under all operating conditions, DRM19-opt agrees well with the experimental data, with the average error decreasing from 17.33% to 5.33%. Note that the experimental data for IDT at p = 3 atm and LFS at p = 0.1 MPa, T = 300 K and p = 0.5 MPa, T = 443 K were not used for training, but the optimized model was still able to predict their IDT and LFS very well. This indicates that the optimized model obtained through GNN has a certain generalization ability.

[0155] 3. Conclusion

[0156] Considering the need for high-fidelity simplified mechanisms in CFD simulations, this paper designs a mechanism optimization strategy based on GNN. The corresponding program code was developed using the PyTorch framework to implement the optimization process.

[0157] Taking the simplified methane combustion mechanism DRM19 as an example, 17 key reactions that have a significant impact on IDT and LFS were selected, and their corresponding pre-exponential factors were optimized. Finally, the optimized model DRM19-opt reduced the MAE of IDT from 32.80% to 14.41% and the MAE of LFS from 17.33% to 5.33%.

[0158] Furthermore, the simulation results of the optimized model show good agreement with the experimental values ​​independent of the training data, demonstrating the effectiveness of GNN. The current method is both simple and efficient, requiring only the establishment of the relationship between kinetic parameters and combustion characteristics through DNN. A novel mechanism optimization method is proposed by using a simple generative network for optimization.

Claims

1. A combustion mechanism optimization method based on generative neural networks, characterized in that, Includes the following steps: Step S1: Optimize the mechanistic dynamic parameters according to the method; Step S2: Based on the optimized mechanism dynamic parameters, analyze the sensitivity, construct the DNN, and optimize the DRM19 mechanism; Step S3: Based on the results of sensitivity analysis, DNN construction, and DRM19 mechanism optimization, draw conclusions.

2. The combustion mechanism optimization method based on generative neural networks according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Perform sensitivity analysis; Step S12: Based on the sensitivity analysis results, establish a deep neural network; Step S13: Generative network based on deep neural network.

3. The combustion mechanism optimization method based on generative neural networks according to claim 1, characterized in that, The specific conclusion of step S3 is as follows: Using the PyTorch framework, corresponding program code was developed to implement the optimization process; The simplified mechanism of methane combustion, DRM19, was selected, and 17 key reactions affecting IDT and LFS were identified. The corresponding pre-exponential factors were optimized. The optimized model DRM19-opt reduced the MAE of IDT from 32.80% to 14.41% and the MAE of LFS from 17.33% to 5.33%. The simulation results of the optimized model are in good agreement with the experimental values ​​independent of the training data, proving the effectiveness of GNN.

4. The combustion mechanism optimization method based on generative neural networks according to claim 2, characterized in that, Step S11 specifically involves: In order to identify the key reactions affecting IDT during combustion, a local sensitivity analysis of the mechanism was performed under experimental temperature, pressure, and equivalence ratio conditions. The calculation formula is as follows: ， Where τ represents IDT, Representing the The reaction rate of one reaction, 2 The reaction rate is doubled; a negative value of S indicates that the corresponding reaction promotes the ignition process. The sensitivity coefficient of LFS is calculated using the normalized sensitivity method in Cantera. The formula is as follows: ， in, Represents LFS, Representative When there are minor changes The changes.

5. The combustion mechanism optimization method based on generative neural networks according to claim 2, characterized in that, Step S12 includes the following steps: Step S121: Define the input vector of the DNN as x = {Φ, p, T, Ai}, where Ai represents the exponential factor of the i-th reaction to be optimized; Step S122: Standardize the dataset input using Z-scores to make the data dimensionless and to unify the magnitudes of different variables. Specifically, the Z-scores are: ， Where, x, x mean x std These represent the original data, mean, and standard deviation, respectively. In this work, the output target value y is determined by IDT and LFS. Step S123: Train the dataset using a neural network with L layers, where the calibrated linear unit is used as the activation function and the mean absolute error is used as the loss function. The formula for calculating the loss function is: ， in, This represents the predicted value of the DNN. This represents the corresponding actual value.

6. The combustion mechanism optimization method based on generative neural networks according to claim 2, characterized in that, Step S13 specifically involves: By coupling with a pre-trained DNN, combustion characteristics IDT and LFS corresponding to kinetic parameters are quickly obtained, and a custom GNN loss function is used to guide the generation of better parameters. The loss function of the GNN is expressed as: ， in, and These are hyperparameters representing the importance of IDT and LFS. and These are simulated values ​​obtained through a DNN. and These are experimental values.

7. The combustion mechanism optimization method based on generative neural networks according to claim 1, characterized in that, The sensitivity analysis in step S2 includes the following steps: Step S211: Local sensitivity analysis was performed under the conditions of Φ = 2.0, T = 1600 K, p = 1, 5, and 10 atm; Step S212: Sensitivity analysis of LFS was performed at p = 0.1 MPa, Φ = 0.7, T = 373 and 443 K.

8. The combustion mechanism optimization method based on generative neural networks according to claim 1, characterized in that, The construction of the DNN in step S2 includes the following steps: Step S221: According to the experimental working conditions of IDT, the experimental conditions are: Φ∈[0.5,2.0], p∈[1 atm,10 atm], T∈[1300 K,2000 K], sampling is performed within the range of variation of the pre-exponential factors of 17 key reactions; Step S222: Simulate all inputs using Cantera to obtain the corresponding IDT and LFS; Step S223: The input data and simulation results constitute a dataset, which is divided into a training set, a validation set, and a test set in a ratio of 0.7:0.2:0.

1.

9. The combustion mechanism optimization method based on generative neural networks according to claim 1, characterized in that, The optimized DRM19 mechanism in step S2 includes the following steps: Step S231: Based on the trained DNN, GNN is used to optimize the parameters of 17 key reactions in the DRM19 dynamic model; Step S232: During training, the GNN receives a set of random 17-dimensional noise and generates a set of 17-dimensional data as potential optimization parameters. Step S233: After combining these generated parameters with experimental conditions, use a pre-trained DNN to quickly predict the IDT and LFS of the dynamic model corresponding to these parameters; Step S234: Calculate the loss using a custom GNN loss function and provide feedback to the GNN to guide it in generating better parameters; Step S235: Place λ IDT and λ LFS The value is adjusted within the range of 0.2 to 0.8; Step S236: After multiple tests, the results show that when λ IDT = 0.7 and λ LFS The optimized dynamic model performs best when the value is 0.3.

Images