A deep neural network-based combustion chemical reaction calculation acceleration method

Abstract

The application discloses a combustion chemical reaction calculation acceleration method based on a deep neural network, relates to the field of combustion numerical simulation, and comprises the following steps: global manifold sampling, which is carried out in a predetermined thermochemical phase space; preprocessing of the sampling data, which comprises Box-Cox transformation and zero-mean standardization; training and saving of a deep neural network model; coupling of the pre-trained deep neural network model with a computational fluid dynamics program to realize combustion numerical simulation calculation. Through the chemical reaction replacement model based on the deep neural network, the strong rigidity in the combustion chemical reaction is removed, the chemical reaction rigid differential equation set is solved at a larger time step, and the chemical reaction multi-grid parallel calculation can be realized through a GPU, so that the calculation efficiency is greatly improved under the premise of ensuring the accuracy, and the method is efficient and convenient.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of combustion numerical simulation, in particular to a combustion chemical reaction calculation acceleration method based on a deep neural network. BACKGROUND

[0002] Aero-engine, scramjet engine and internal combustion engine are key equipment power for national defense and national economy, and numerical simulation is a key tool for scientific research and industrial design. Combustion is a typical multi-scale problem, and the difficulty of numerical simulation lies in the coupling of complex chemical reactions and flow of actual fuel. In the combustion numerical simulation involving detailed chemical mechanism, the number of components is large, the dimension is high, and the chemical reaction is rigid. Solving chemical reactions usually accounts for 70%-90% of the calculation time, and accelerating chemical reaction operation is a bottleneck that needs to be broken through to realize efficient numerical simulation.

[0003] In the numerical simulation of combustion of hydrocarbon fuel, solving chemical reactions accounts for most of the calculation time. Because the combustion chemistry is rigid, conventional numerical simulation needs a time step much smaller than the Courant number to evolve the reaction source term. The combustion chemical source term replacement model based on machine learning algorithm is not constrained by chemical rigidity, and can predict the chemical reaction term with a larger time step. Compared with various tabulation methods, it has the advantage of small memory overhead. However, the existing neural network replacement model has a large constraint on the training data, and is severely dependent on specific problem scenarios. A large number of numerical simulations of similar scenarios are needed to obtain the required samples for training. Additional perturbation methods are often needed to enhance the samples, and the overall dependence on specific problems is heavy.

[0004] Therefore, the skilled in the art is committed to developing a combustion chemical reaction calculation acceleration method based on a deep neural network. SUMMARY

[0005] In view of the above defects of the prior art, the technical problem to be solved by the present application is to remove the strong rigidity in combustion chemical reactions, realize multi-grid chemical reaction parallel computing, and effectively realize chemical reaction calculation acceleration in the numerical simulation of combustion of hydrocarbon fuel under the premise of ensuring accuracy.

[0006] To achieve the above purpose, the present application provides a combustion chemical reaction calculation acceleration method based on a deep neural network, characterized in that the method comprises the following steps:

[0007] S101: Global manifold sampling, the sampling is performed in a predetermined thermo-chemical phase space;

[0008] S103: Preprocessing of the sampling data, the preprocessing includes Box-Cox transformation and zero-mean standardization;

[0009] S105: training and saving the deep neural network model;

[0010] S107: coupling the pre-trained deep neural network model with a computational fluid dynamics program to realize numerical simulation calculation of combustion.

[0011] Further, in the step S101, the thermo-chemical phase space is a vector space composed of thermo-chemical states, and the thermo-chemical states include temperature, pressure, and mass fraction of each component.

[0012] Further, the components in the thermo-chemical phase space include methane, oxygen, and nitrogen, the mass fraction of each component is 1:1:3.76, the initial temperature in the thermo-chemical phase space is 300K, and the initial pressure in the thermo-chemical phase space is 1 atm.

[0013] Further, the global manifold sampling includes the following steps:

[0014] S1011: performing zero-dimensional manifold sampling on a low-dimensional sub-manifold in the thermo-chemical space to obtain a zero-dimensional manifold data set;

[0015] S1012: randomly collecting a large number of samples in the thermo-chemical phase space according to a logarithmic scale to obtain a multi-scale sampling data set;

[0016] S1013: taking the boundary of the rate of change of each dimension in the zero-dimensional manifold data set as a reference, removing samples in the multi-scale sampling data set whose time change rate exceeds the boundary range, and screening to obtain sample data conforming to actual physical properties;

[0017] S1014: performing evolution, generating labels, and merging processing on the sample data to generate a data set; the evolution generates state quantities after 1E-3s time steps from predetermined current state quantities, the generating labels generates state quantities after 1E-6s time steps from predetermined current state quantities; the merging includes merging the data obtained by evolution as input data and merging the data obtained by generating labels as target data.

[0018] Further, in the step S103, the Box-Cox transformation adopts the following transformation method:

[0019]

[0020] wherein x is an independent variable, and λ is a variable hyperparameter.

[0021] Further, in the step S103, the zero-mean standardization means that all feature values of input samples are subtracted by the mean of the samples and then divided by the standard deviation of the samples, so that the mean of the samples is zero and the variance is 1.

[0022] Further, in the step S105, the deep neural network is a full connection deep neural network, the training data adopts the generated data set in the global manifold sampling, and the network model parameter of the deep neural network after training is saved as a binary format file.

[0023] Further, in the step S107, the computational fluid dynamics program includes EBI-DNS, and the chemical reaction calculation part of the computational fluid dynamics program is replaced by using the deep neural network model.

[0024] Further, the neural network model supports deployment on a CPU and / or a GPU, supports parallel computing, the calculation domain is a matrix area of 1.5 cm*1.5 cm, and the calculation grid is N=512*512.

[0025] Further, the ignition source radius is 0.02 cm, the ignition source energy in the thermo-chemical phase space is w / m3, the ignition lasts for 0.2 ms, and the initial velocity field and the initial vorticity field are generated by the following method.

[0026]

[0027] wherein E(k) is an energy spectrum, k is a frequency, the average velocity u rms = 3 m / s, the constant k e = 418.67.

[0028] In the preferred embodiment of the present application, the present application has the following beneficial effects relative to the prior art:

[0029] 1. The present application removes the strong rigidity in the combustion chemical reaction by using the chemical reaction replacement model based on the deep neural network, realizes solving the chemical reaction rigid differential equation set with a larger time step, and can realize the chemical reaction multi-grid parallel computing through the GPU, thereby greatly improving the calculation efficiency under the premise of ensuring the accuracy.

[0030] 2. The chemical reaction replacement model based on the neural network proposed in the present application can be used for laminar flow, jet flame and turbulent flame numerical simulation without fine tuning, that is, plug and play, and is efficient and convenient.

[0031] 3. The global manifold sampling proposed in the present application randomly samples the entire chemical phase space and does not depend on the specific example chemical reaction prior setting and flow field information.

[0032] The concept, specific structure and generated technical effects of the present application will be further described below with reference to the accompanying drawings, so as to fully understand the purpose, features and effects of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a flow chart of a deep neural network based combustion chemical reaction calculation acceleration method of a preferred embodiment of the present application;

[0034] Figure 2 is an initial velocity field schematic diagram of a deep neural network based combustion chemical reaction calculation acceleration method of a preferred embodiment of the present application;

[0035] Figure 3 is an initial vorticity field schematic diagram of a deep neural network based combustion chemical reaction calculation acceleration method of a preferred embodiment of the present application;

[0036] Figure 4 is a turbulent flame temperature cloud atlas schematic diagram after 1ms of propulsion of a deep neural network based combustion chemical reaction calculation acceleration method of a preferred embodiment of the present application;

[0037] Figure 5 is a turbulent flame temperature cloud atlas schematic diagram after 1ms of propulsion of a deep neural network based combustion chemical reaction calculation acceleration method of a preferred embodiment of the present application;

[0038] Figure 6 is a calculation time consumption schematic diagram after 1ms of propulsion of a deep neural network based combustion chemical reaction calculation acceleration method of a preferred embodiment of the present application. DETAILED DESCRIPTION

[0039] The following reference to the drawings introduces the preferred embodiments of the present application, making the technical content of the present application clearer and easier to understand. The present application can be embodied in many different forms of embodiments, and the protection scope of the present application is not limited to the embodiments mentioned in the text.

[0040] In the drawings, components of the same structure are denoted by the same reference numerals, and components having similar structures or functions are denoted by similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present application is not limited to the size and thickness of each component. In order to make the drawing clearer, the thickness of the components is appropriately exaggerated in some places in the drawing.

[0041] As shown in Figure 1 , a deep neural network based combustion chemical reaction calculation acceleration method provided by an embodiment of the present application includes the following steps:

[0042] S101: Global manifold sampling, the sampling is performed in a predetermined thermochemical phase space.

[0043] The thermo-chemical phase space is a vector space composed of thermo-chemical states, including temperature, pressure and mass fraction of each component, wherein the components include methane, oxygen and nitrogen, the mass fraction of each component is 1:1:3.76, the initial temperature in the thermo-chemical phase space is 300K, and the initial pressure in the thermo-chemical phase space is 1 atm. The ignition source radius in the thermo-chemical phase space is 0.02 cm, the ignition source energy is w / m3, the ignition lasts for 0.2 ms, and the initial velocity field and the initial vorticity field are generated by the following method:

[0044]

[0045] wherein E(k) is an energy spectrum, k is frequency, the average velocity u rms = 3 m / s, and the constant k e = 418.67.

[0046] When performing global manifold sampling, the following steps are included:

[0047] S1011: performing zero-dimensional manifold sampling on a low-dimensional sub-manifold in the thermo-chemical space to obtain a zero-dimensional manifold data set;

[0048] S1012: randomly collecting a large number of samples in the thermo-chemical phase space according to a logarithmic scale to obtain a multi-scale sampling data set;

[0049] S1013: taking the boundaries of the rate of change of each dimension in the zero-dimensional manifold data set as a reference, removing samples in the multi-scale sampling data set whose time change rate exceeds the boundary range, and screening to obtain sample data conforming to actual physical properties;

[0050] S1014: performing evolution, generating labels and merging processing on the sample data to generate a data set; the evolution generates state quantities after 1E-3s time steps from predetermined current state quantities, the generating labels generate state quantities after 1E-6s time steps from predetermined current state quantities; the merging includes merging the data obtained by evolution as input data and merging the data obtained by generating labels as target data.

[0051] S103: performing preprocessing on the sampling data, including Box-Cox transformation and zero-mean standardization.

[0052] The Box-Cox transformation adopts the following transformation method:

[0053]

[0054] wherein x is an independent variable, and λ is a variable hyperparameter.

[0055] In the zero-mean standardization process, all feature values of the input sample are subtracted by the mean value of the sample and then divided by the standard deviation of the sample, so that the mean value of the sample is zero and the variance is 1.

[0056] S105: training and saving the deep neural network model.

[0057] The deep neural network is a fully connected deep neural network, the training data uses the data set generated in the global manifold sampling, and the network model parameters of the deep neural network after training are saved as a binary format file for direct import and use in the future.

[0058] S107: coupling the pre-trained deep neural network model with the computational fluid dynamics program to realize combustion numerical simulation calculation.

[0059] In the specific combustion numerical simulation calculation, the pre-trained deep neural network model is coupled with the computational fluid dynamics program, such as EBI-DNS, and the deep neural network model is used to replace the chemical reaction calculation part of the computational fluid dynamics program, the neural network model supports deployment on CPU and / or GPU and supports parallel computing, the calculation domain uses a 1.5cm*1.5cm matrix area, and the calculation grid is N=512*512.

[0060] The present application removes the strong rigidity in the combustion chemical reaction, and the single model can be used for laminar flow, jet flame and turbulent flame numerical simulation without fine tuning, can realize multi-grid parallel computing of chemical reaction through GPU, can solve chemical reaction rigid differential equations with large time step, can realize multi-grid chemical reaction parallel computing, can effectively realize chemical reaction calculation acceleration in combustion numerical simulation of hydrocarbon fuel under the premise of ensuring accuracy, greatly improves the calculation efficiency, and can be used immediately, and is efficient and convenient.

[0061] In addition, the global manifold sampling proposed in the present application randomly samples the entire chemical phase space and does not depend on specific prior settings of chemical reactions and flow field information of examples; the neural network algorithm proposed can be applied to multiple combustion test examples and has the advantages of good universality, high accuracy and the like.

[0062] The present application will be described in detail below in combination with the preferred embodiments of the present application.

[0063] As Figure 2As shown, the preferred embodiment of the present application relates to a neural network-based combustion chemical reaction calculation acceleration method, including: the calculation domain is a matrix area of 1.5cm*1.5cm, and the calculation grid is N=512*512. Common hydrocarbon fuels such as methane, ethylene, etc., the chemical reaction mechanism selects the methane DRM19 mechanism of 21 components and 84 reactions commonly used in numerical simulation of methane combustion. In the initial state, the calculation area is filled with a mixture of methane, oxygen and nitrogen gas, and the molar ratio is 1:1:3.76. The initial temperature is 300K, the initial pressure is 1atm, the coordinate system is established with the center of the reaction container as the origin, the ignition source is located at (0.75cm, 0.75cm), the radius is 0.02cm, the ignition source energy is w / m3, and the ignition lasts for 0.2ms. The initial velocity field and the initial vorticity field are generated by the Passot-Pouquet isotropic kinetic energy spectrum, specifically:

[0064]

[0065] Wherein, E(k) is the energy spectrum, k is the frequency, u rms The average velocity u rms =3m / s, the constant k e =418.67.

[0066] The initial velocity field is as shown in Figure 3 The initial vorticity field is as shown in Figure 4 The turbulent flame temperature cloud chart after 1ms is as shown in Figure 5

[0067] The deep neural network-based combustion chemical reaction surrogate model established in the preferred embodiment includes the following steps:

[0068] Step 1, given the temperature and pressure range, perform global manifold sampling in the thermochemical phase space;

[0069] Step 2, combine Box-Cox transformation and zero-mean standardization to preprocess the sampling data;

[0070] Step 3, train and save the deep neural network model;

[0071] Step 4, the pre-trained deep neural network model is coupled with the computational fluid dynamics (CFD) program, the chemical reaction part is calculated by the neural network model, and the combustion is simulated.

[0072] ​This embodiment compares the computation time required to calculate chemical reactions in a combustion numerical simulation using a neural network model-based method and the existing classical differential equation solver CVODE. Specifically, while ensuring the accuracy of the numerical simulation, both methods are used as solvers for the chemical reactions. From the start of the combustion simulation ignition to 1 ms after the operation, the computation time of the neural network model on the CPU, the computation time of the neural network model on the GPU, and the computation time of CVODE on the CPU are compared in the overall simulation.

[0073] The thermochemical phase space in the preferred embodiment described above is a vector space composed of thermochemical states, including temperature, pressure, and the mass fraction of each component. Global manifold sampling, by providing an initial state space based on temperature, pressure, and equivalence ratio, evolves its initial state vector to a chemically stable state, forming a low-dimensional submanifold in the thermochemical space. This includes the following specific steps: First, the method of randomly selecting data on this low-dimensional manifold is called zero-dimensional manifold sampling. The dataset obtained through sampling is a zero-dimensional manifold dataset, denoted as D. MF Secondly, multi-scale sampling is performed throughout the entire thermochemical phase space, that is, a large number of samples are randomly collected according to a logarithmic scale to obtain a multi-scale sampling dataset, denoted as D. MS Then, using the manifold dataset D... MF Using the boundaries of the rate of change in each dimension as a benchmark, the multi-scale dataset D was removed. MS Samples whose rate of change over time exceeds the boundary range are selected, thereby identifying a batch of samples that conform to actual physical properties.

[0074] Regarding the evolution in the above steps, given the current state, the state after a 1E-3s time step is generated using the open-source chemical kinetics software Cantera. For generating labels, given the current state, the state after a 1E-6s time step is generated using the open-source chemical kinetics software Cantera. The data obtained by merging the evolution is the input data, and the data obtained by merging the labels is the target data.

[0075] When performing data preprocessing, the following transformation method is used for the Box-Cox transformation:

[0076]

[0077] Where x is the independent variable and λ is a variable hyperparameter.

[0078] Data preprocessing also includes zero-mean standardization, which involves subtracting the mean of the sample from all the feature values ​​of the input sample and dividing by the standard deviation of the sample, so that the mean of the sample is zero and the variance is 1.

[0079] In training the deep neural network, a deep learning open source framework pytorch is used to train a fully connected deep neural network, and the training data uses the data set of the above merging step, and the network model parameters after training are saved in binary format file for subsequent use in practical application.

[0080] In coupling the deep neural network model with the widely used computational fluid dynamics program, based on the existing mature computational fluid dynamics program EBI-DNS in the field of combustion, the chemical reaction calculation part is changed to deep neural network operation, and the deep neural network coupled fluid is realized to realize the numerical simulation of combustion.

[0081] The chemical reaction mechanism selects the base component, base reaction and reaction rate coefficient A / B / Ea in the methane DRM19 mechanism of 21 components and 84 reactions commonly used in numerical simulation of methane combustion as shown in Table 1.

[0082] Table 1 Base component and base reaction in 21 components and 84 reactions

[0083]

[0084]

[0085]

[0086]

[0087] The deep neural network-based general replacement model for chemical reactions proposed in the preferred embodiment removes the strong rigidity in combustion chemical reactions, and a single model can be used for laminar flow, jet flame and turbulent flame numerical simulation without fine tuning. The GPU can realize multi-grid parallel computing of chemical reactions, and can realize multi-grid parallel computing of chemical reactions. The GPU model involved in the preferred embodiment is NVIDIA Tesla V100-SXM2-32GB, and the CPU model is Intel(R) Xeon(R) Platinum8260 CPU@2.40GHz. The test will compare the acceleration effect of CPU-GPU parallel computing and only CPU parallel computing. The calculation time after 1ms of the preferred embodiment is as follows. Figure 6As shown, the chemical reactions in the combustion simulation are calculated using CVODE and deep neural network (DNN) respectively, DNN-CPU refers to the operation of the DNN model on the CPU, and DNN-GPU refers to the operation of the DNN model on the GPU. The preferred embodiment shows that, under the premise of ensuring calculation accuracy, using a neural network model combined with a CPU can achieve about 10 times acceleration of chemical reaction calculation, and combined with GPU parallel calculation can achieve about 30 times calculation ratio. The chemical reaction acceleration algorithm based on the neural network model can significantly improve the calculation efficiency, effectively realize the chemical reaction calculation acceleration in the combustion numerical simulation of hydrocarbon fuel under the premise of ensuring the accuracy, greatly improve the calculation efficiency, and be plug and play, efficient and convenient.

[0088] The preferred embodiments of the present application are described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and changes without creative labor based on the concept of the present application. Therefore, any technical solutions obtained by logical analysis, reasoning or limited experiments based on the prior art according to the concept of the present application shall be within the protection scope defined by the claims.

Claims

1. A method for accelerating the computation of combustion chemical reactions based on deep neural networks, characterized in that, The method includes the following steps: S101: Global manifold sampling, the sampling being performed within a predetermined thermochemical phase space; S103: Preprocess the sampled data, the preprocessing including Box-Cox transformation and zero-mean standardization; S105: Train and save the deep neural network model; S107: Couple the pre-trained deep neural network model with the computational fluid dynamics program to achieve numerical simulation calculation of combustion; in, The global manifold sampling includes the following steps: S1011: Perform zero-dimensional manifold sampling on the low-dimensional submanifolds in the thermochemical phase space to obtain a zero-dimensional manifold dataset; S1012: Randomly collect a large number of samples in the thermochemical phase space according to a logarithmic scale to obtain a multi-scale sampling dataset; S1013: Using the boundaries of the rate of change of each dimension in the zero-dimensional manifold dataset as a benchmark, remove samples in the multi-scale sampling dataset whose rate of change of time exceeds the boundary range, and filter to obtain sample data that conforms to the actual physical properties. S1014: The sample data is subjected to evolution, label generation, and merging processes to generate a dataset; the evolution generates a state value 1E-3 s after a predetermined current state value, and the label generation generates a state value 1E-6 s after a predetermined current state value; the merging process includes merging the data obtained from the evolution process as input data and merging the data obtained from the label generation process as target data. In step S107, the computational fluid dynamics program includes EBI-DNS, which replaces the chemical reaction calculation part of the computational fluid dynamics program with the deep neural network model.

2. The method as described in claim 1, characterized in that, In step S101, the thermochemical phase space is a vector space composed of thermochemical states, which include temperature, pressure and mass fraction of each component.

3. The method as described in claim 2, characterized in that, The components in the thermochemical phase space include methane, oxygen, and nitrogen, with a mass fraction of 1:1:3.

76. The initial temperature in the thermochemical phase space is 300 K, and the initial pressure in the thermochemical phase space is 1 atm.

4. The method as described in claim 1, characterized in that, In step S103, the Box-Cox transformation employs the following transformation method: ， Where x is the independent variable and λ is a variable hyperparameter.

5. The method as described in claim 4, characterized in that, In step S103, zero-mean standardization means subtracting the mean of the sample from all feature values ​​of the input sample and then dividing by the standard deviation of the sample, so that the mean of the sample is zero and the variance is 1.

6. The method as described in claim 1, characterized in that, In step S105, the deep neural network is a fully connected deep neural network, the training data is the dataset generated in the global manifold sampling, and after training, the network model parameters of the deep neural network are saved as a binary format file.

7. The method as described in claim 1, characterized in that, The neural network model supports deployment on CPU and / or GPU and supports parallel computing. The computing domain is a rectangular area of ​​1.5cm × 1.5cm, and the computing grid is N = 512 × 512.

8. The method as described in claim 7, characterized in that, The radius of the ignition source within the computational domain is 0.02 cm, and the ignition source energy is 1 × 10⁻⁶. 11 w / m3, ignition duration 0.2ms, initial velocity field and initial vorticity field are generated as follows: ， in, It is an energy spectrum. It refers to frequency and average speed. =3m / s, constant =418.67.

Images