Intelligent solving method for supersonic combustion flow field based on hierarchical physical constraints

Abstract

The application discloses a kind of based on layered physical constraint supersonic combustion flow field intelligent solving method, it is related to supersonic combustion chamber combustion simulation calculation and deep learning cross field, comprising: S1, the simulation data reliability obtained by numerical simulation of supersonic combustion flow is verified;S2, three-dimensional combustion flow field data set is constructed based on the final turbulent flow model parameter obtained in S1;S3, according to the generation principle of scramjet engine combustion flow, construct physical mechanism model;S4, according to three-dimensional combustion flow field data set, construct neural network model;S5, the prediction results of each layer of neural network model are constrained in stages using physical mechanism model, to obtain three-dimensional flow field intelligent solving model.The application is guaranteed while high-precision generation of three-dimensional supersonic combustion flow evolution process, further accelerates the generation speed of model, for accelerating the numerical simulation process of scramjet, promote the optimization process of scramjet engine provides technical support.

Description

TECHNICAL FIELD

[0001] The present application relates to the cross field of supersonic combustion chamber combustion simulation calculation and deep learning. More specifically, the present application relates to a layered physical constraint-based intelligent solution method for supersonic combustion flow field. BACKGROUND

[0002] As an ideal power device for modern supersonic aircraft, the scramjet engine has many advantages such as light weight, high speed, high unit thrust, etc., and is one of the key technologies for countries to occupy the advantage of space and time. The airflow speed in the scramjet engine combustion chamber is on the order of kilometers per second, which is similar to lighting a match in a tornado, bringing problems such as mixing, ignition, flame stabilization and propagation to combustion organization. Under supersonic conditions, oblique shock waves and expansion waves are reflected multiple times between the combustion chamber wall and the fuel flow, and pass through the middle of the airflow, so that the combustion flow characteristics inside the combustion chamber have strong shock boundary layer interference, shock and combustion area coupling, and other strong nonlinear characteristics, which significantly affect the engine operating state.

[0003] Numerical simulation calculation is an important means to study supersonic aircraft, including direct numerical calculation (DNS), finite difference, finite element method and meshless numerical calculation method, etc. They provide important technical support for how engineering systems combine with fluid flow and combustion state by numerically solving Navier-Stokes equations (NS equations) in time and space. However, the traditional numerical simulation software based on equation iteration solution cannot fully meet the large-scale engineering application with hysteresis characteristics, especially for the modeling of strong swirling flow, three-dimensional unsteady two-phase multi-physical strong coupling process in the combustion chamber and high-precision thermochemical reaction kinetics solution, etc. The calculation efficiency and accuracy of existing numerical simulation software are severely limited. Therefore, the study of the alternative method of layered physical constraint-based supersonic combustion flow numerical simulation calculation is helpful to replace the traditional iterative solution calculation based on given inflow conditions, to realize the "fast" and "accurate" direct simulation of supersonic combustion flow, i.e. visualization, which has important theoretical and engineering significance for improving the design and analysis ability in the field of aviation and aerospace, accelerating technological innovation, and reducing research and development costs, etc. SUMMARY

[0004] An object of the present application is to solve at least the above problems and / or deficiencies, and to provide at least the advantages described later.

[0005] To achieve these objects and other advantages and in view of its purposes, a layered physical constraint-based intelligent solution method for supersonic combustion flow field is provided, comprising:

[0006] S1. Compare the experimental data obtained from the ground pulse combustion wind tunnel test with the simulation data obtained from the numerical simulation of supersonic combustion flow, and adjust the turbulence model parameters based on the comparison results before re-performing the numerical simulation until the simulation data matches the experimental data.

[0007] S2. Based on the final turbulence model parameters obtained in S1, perform numerical simulation calculations of supersonic combustion flow under different incoming flow conditions to construct a three-dimensional combustion flow field dataset.

[0008] S3. Based on the generation principle of combustion flow in scramjet engines, construct a physical mechanism model;

[0009] S4. Based on the three-dimensional combustion flow field dataset, construct a neural network model based on hierarchical encoding and fusion decoding;

[0010] S5. A physical mechanism model is used to constrain the prediction results of each layer of the neural network model in stages, so as to obtain a three-dimensional flow field intelligent solution model based on hierarchical physical information constraints.

[0011] S6. Apply the three-dimensional flow field intelligent solution model to the prediction of three-dimensional supersonic combustion flow.

[0012] Preferably, in S1, the method for adjusting the turbulence model parameters is as follows:

[0013] S11. Based on the ground pulse combustion wind tunnel test, pressure sensors are used to obtain test data of combustion chamber wall pressure.

[0014] S12. Based on the given parameters of the scramjet engine, use numerical simulation software to obtain simulation data of the entire three-dimensional space of the scramjet engine.

[0015] S13. Based on the arrangement of pressure sensors in the ground pulse combustion wind tunnel test, select simulated wall pressure data from the simulation data at the same location as the scramjet engine.

[0016] S14. Judge the consistency between the wall pressure test data and the wall pressure simulation data. If the judgment result is inconsistent, adjust the turbulence model parameters and return to S12. Otherwise, use the current turbulence model parameters as the final turbulence model parameters.

[0017] Preferably, in S2, the three-dimensional combustion flow field dataset is constructed as follows:

[0018] Three-dimensional numerical simulation calculations of scramjet engine combustion were performed based on the final turbulence model parameters. According to the wave system structure and flame distribution law of the multi-physics combustion field of scramjet engine, the sampling points in the three-dimensional flow field were obtained by adopting the segmented random sampling and local dense sampling strategies.

[0019] The x, y, z coordinates of the sampling points, non-stationary time, total incoming temperature, total pressure, Mach number, hydrogen injection location, and injection flow rate are used as input data in the three-dimensional combustion flow field dataset.

[0020] The velocity field, pressure field, Mach number field, total temperature, static temperature, and concentration field of each component in the x, y, and z directions in three-dimensional space are used as output data in the three-dimensional combustion flow field dataset.

[0021] Preferably, the segmented random sampling and local dense sampling strategy refers to: dividing the entire three-dimensional physical field into multiple sub-regions, determining the wave system structure and combustion complexity in the corresponding sub-region based on the density gradient changes of each sub-region, and setting different numbers of sampling points according to the complexity of the corresponding sub-region.

[0022] Preferably, in S3, the physical mechanism model includes: a physical model PY 1 and physical model PY 2;

[0023] Among them, physical model PY 1 is represented as:

[0024]

[0025] Physical Model PY 2 is represented as:

[0026]

[0027] In the above formula, p Represents the density field. u , v , w They represent in x direction, y direction, z velocity field in the direction, p Represents a pressure field. express x Shear stress in a plane express y Shear stress in a plane express z Shear stress in a plane express x plane and y Shear stress between planes express x plane and z Shear stress between planes express y plane and z Shear stress between planes Indicates the incoming Reynolds number, com Indicates the number of components. Ci Represents the concentration field of each component. i The first fuel i Number of components Indicates the total temperature. Indicates static temperature. r Indicates specific heat ratio. Mach This indicates the Mach number field.

[0028] Preferably, in S4, the fitting process of the neural network model is characterized by the following formula:

[0029]

[0030] In the formula, This indicates the fitting result. Represents a neural network model. x , y , z The values ​​represent the sampling point coordinates, t represents the non-stationary time, T0 represents the total incoming flow temperature, P0 represents the total pressure, Ma0 represents the Mach number, K represents the hydrogen injection location, and L represents the injection flow rate. u , v , w They represent in x direction, y direction, z velocity field in the direction, Ci Represents the concentration field of each component. Indicates the total temperature. Indicates static temperature. Mach Indicates Mach field.

[0031] Preferably, in S4, the neural network model includes:

[0032] Parallel configuration of fully connected neural networks FC1 and FC2;

[0033] The first and second layered convolutional modules are respectively set on the output side of the fully connected neural network FC1 and the fully connected neural network FC2;

[0034] A feature decoding module that fuses the outputs of the first and second layered convolutional modules;

[0035] Among them, on the output side of the fully connected neural network FC1, there is also a dynamic fusion module that is set in parallel with the first layered convolution module to adaptively and dynamically transform the nonlinear mapping of the fully connected neural network FC1. The nonlinear mapping data of the fully connected neural network FC1 comes from the inflow total temperature T0, total pressure P0, Mach number Ma0, hydrogen injection position k, and injection flow rate L in the three-dimensional combustion flow field dataset.

[0036] At the front end of the fully connected neural network FC2, there is also a position encoding module that performs polar coordinate encoding of the coordinates x, y, z of the sampling points in the three-dimensional combustion flow field dataset.

[0037] Preferably, the feature decoding module includes:

[0038] Conv5 is a convolutional layer that performs convolution processing on the fused output of the first layer convolutional module and the dynamic fusion module;

[0039] Convolutional layers Conv6 and Conv7 that group the output of convolutional layer Conv5;

[0040] The output of the second layered convolutional module is sequentially processed by convolutional layers Conv8 and Conv9, and then by an upscale layer that performs upsampling on the output of Conv9.

[0041] After pixel-wise addition and fusion processing of the outputs of convolutional layers Conv6, Conv7, and Upscale, the fused feature information is output by convolutional layer Conv10 using multi-physics field.

[0042] Preferably, in S5, the phased constraint refers to:

[0043] In the first stage, a neural network model was used to analyze the velocity and pressure fields in the x, y, and z directions. p The Mach number field is used for prediction, and the physical model in the physical mechanism model is used. PY 1. Constrain the prediction results;

[0044] In the second stage, a neural network model was used to analyze the total temperature. , static temperature and the concentration field of each component Ci To make predictions, a physical model from the physical mechanism model is used. PY 2. Constrain the prediction results.

[0045] Preferably, in S5, the loss function in the three-dimensional flow field intelligent solution model is... loss It is characterized by the following formula:

[0046] ;

[0047] In the above formula, For physical models PY A deviation of 1, and ; For physical models PY The deviation of 2, and ; Indicates data-driven error, and It is characterized by the following formula:

[0048] ;

[0049] in, This represents labeled data, where m represents the total number of samples. This indicates the fitting result. u , v , w They represent in x direction, y direction, z velocity field in the direction, Ci Represents the concentration field of each component. Indicates the total temperature. Indicates static temperature. Mach Indicates Mach number. p Represents the density field. p It represents a pressure field.

[0050] The present invention has at least the following beneficial effects:

[0051] In the numerical simulation of supersonic combustion chambers, the reliability of the numerical simulation results of the three-dimensional turbulent combustion process is verified by ground wind tunnel test results. Then, a high-quality three-dimensional combustion flow field dataset is constructed, and a three-dimensional flow field intelligent solution model based on hierarchical physical constraints is constructed by combining deep neural networks and combustion mechanisms. While ensuring the high-precision generation of the three-dimensional supersonic combustion flow evolution process, the model generation speed is further accelerated, providing technical support for accelerating the numerical simulation process of scramjet engines and promoting the optimization process of scramjet engines.

[0052] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0053] Figure 1 This is a flowchart of an intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints in an embodiment of the present invention;

[0054] Figure 2This is a schematic diagram of the network structure of the three-dimensional flow field intelligent solution model in an embodiment of the present invention.

[0055] Figure 3 This is a schematic diagram of the network structure of the neural network model in an embodiment of the present invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0057] A hierarchical physical constraint-based intelligent solution method for supersonic combustion flow fields replaces traditional numerical simulation calculations, improving the accuracy and speed of numerical simulation of turbulent combustion in the supersonic combustion chamber of scramjet engines. (Reference) Figure 1 Specifically, it includes the following:

[0058] S1. Based on the turbulence model and ground-based pulse combustion wind tunnel tests, verify the reliability (also known as consistency) of the numerical simulation calculations of supersonic combustion flow. This specifically includes the following steps:

[0059] S11. First, given the total temperature, total pressure, Mach number, hydrogen injection location, injection flow rate, and other parameter data of the scramjet engine; and based on the ground pulse combustion wind tunnel test, use high-precision pressure sensors to obtain the pressure test data of the upper and lower walls of the combustion chamber, and use a high-speed schlieren camera to obtain the density gradient field inside the combustion chamber.

[0060] S12. Using independently controllable numerical simulation software, based on the given parameters of the scramjet engine (i.e., total temperature, total pressure, Mach number, hydrogen injection location, and injection flow rate), the simulation data of the scramjet engine in the entire three-dimensional space is obtained (i.e., velocity fields u, v, and w in the x, y, and z directions, pressure field p, Mach number field, and total temperature field). , static temperature And the concentration field Ci of each component (i represents the number of components in the fuel));

[0061] S13. Based on the arrangement of the pressure sensors in the ground pulse combustion wind tunnel test, select simulated wall pressure data at the same location of the scramjet engine from the simulation data.

[0062] S14. Compare the wall pressure test data with the wall pressure simulation data for consistency. Adjust the turbulence model parameters based on the comparison results (in specific implementation, the turbulence model can be set to the SST kw turbulence model). Return to S12 to re-perform the numerical simulation calculation until the wall pressure simulation data obtained from the numerical simulation calculation is consistent with the wall pressure test data obtained from the ground pulse combustion wind tunnel test. This verifies the reliability of the supersonic combustion flow numerical simulation calculation.

[0063] S2. Based on the reliability verification results, numerical simulation calculations of supersonic combustion flow under different incoming flow conditions are performed according to the adjusted turbulence model parameters to construct a three-dimensional combustion flow field dataset, which specifically includes the following:

[0064] Based on the adjusted turbulence model parameters, a three-dimensional numerical simulation of the scramjet engine combustion was performed. According to the wave system structure and flame distribution law of the multi-physics combustion field of the scramjet engine, a segmented random sampling and local dense sampling strategy was adopted to obtain the distribution of sampling points in the three-dimensional flow field.

[0065] Among them, the segmented random sampling and local dense sampling strategy refers to: dividing the entire three-dimensional physical field into multiple sub-regions, judging the complexity of the wave system structure and combustion in the sub-region based on the density gradient change of the sub-region, and finally setting different numbers of sampling points according to the complexity of the sub-region;

[0066] The x, y, z coordinates of the acquired sampling points, the non-fixed field time t, the total temperature of the incoming flow T0, the total pressure P0, the Mach number Ma0, the hydrogen injection position k, and the injection flow rate L are used as input data in the three-dimensional combustion flow field dataset.

[0067] The obtained velocity fields u, v, w in the x, y, and z directions of the three-dimensional space, the pressure field p, the Mach number field Mach, and the total temperature are... , static temperature The concentration fields of each component Ci (i represents the number of components in the fuel) are used as output data in the three-dimensional combustion flow field dataset.

[0068] S3. Based on the generation principle of combustion flow in scramjet engines, construct a physical mechanism model;

[0069] Specifically, based on the conservation of mass, energy, momentum, and composition, a physical mechanism model is constructed. In order to reduce the changes in enthalpy and entropy caused by complex combustion chemical reactions, the physical model has extremely strong non-convex characteristics. The composition conservation law is constrained according to the law that the concentration of each component remains unchanged during the combustion process, so as to prevent the generation of non-physical solutions.

[0070] The physical mechanism model in this step specifically includes "physical model" PY 1 and physical model PY 2;

[0071] Among them, physical model PY 1 is represented as:

[0072]

[0073] Physical Model PY2 is represented as:

[0074]

[0075] In the above formula, p Represents the density field. u , v , w They represent in x direction, y direction, z velocity field in the direction, p Represents a pressure field. express x Shear stress in a plane express y Shear stress in a plane express z Shear stress in a plane express x plane and y Shear stress between planes express x plane and z Shear stress between planes express y plane and z Shear stress between planes Indicates the incoming Reynolds number, com Indicates the number of components. Ci Represents the concentration field of each component. i The first fuel i Number of components Indicates the total temperature. Indicates static temperature. r Indicates specific heat ratio. Mach This indicates the Mach number field.

[0076] S4. Based on the three-dimensional combustion flow field dataset, construct a neural network model based on encoding and decoding;

[0077] In the 3D combustion flow field dataset constructed in S2, 80% of the data was selected as the training set for model construction, and 20% was used to verify the accuracy of the model. A neural network model was used to extract features from the data to obtain prior feature information.

[0078] Specifically, the role of a neural network model is to establish a chain-like functional relationship and nonlinear fitting characteristics between input and output. The fitting process can be represented as follows:

[0079]

[0080] In the formula, This indicates the fitting result. The neural network model is shown; x, y, z represent the coordinates of the sampling point; t represents the non-stationary time; T0 represents the total temperature of the incoming flow; P0 represents the total pressure; Ma0 represents the Mach number; k represents the injection position of hydrogen; L represents the injection flow rate.

[0081] like Figure 3 As shown, the neural network model includes: a fully connected neural network FC1, a first hierarchical convolutional module, a fully connected neural network FC2, a second hierarchical convolutional module, and a feature decoding module;

[0082] The first layered convolutional module includes: a convolutional Conv1 layer, a Bn1 layer, a ReLU layer, a convolutional Conv2 layer, and a ReLU1 layer connected in sequence, as well as an introduced dynamic fusion module;

[0083] The second layered convolution module includes: convolution Conv3 and convolution Conv4 set in parallel, and a Sigmoid at the output of convolution Conv3 and convolution Conv4;

[0084] The feature decoding module includes Conv5, Conv6, Conv7, Conv8, Conv9, an Upscale layer, and Conv10.

[0085] In practical operation, the input data of the neural network model are sampling input points in three-dimensional space, including: total incoming temperature T0, total pressure P0, Mach number Ma0, hydrogen injection position k, and injection flow rate L. Figure 3 In this context, P1, P2…PN correspond to the aforementioned sampling input points (also known as incoming flow parameters), while drop refers to the Dropout technique used in neural networks for model compression. This technique simplifies the network structure by randomly discarding redundant neurons, indirectly reducing the number of parameters and computational overhead, and improving the model's generalization ability. These sampling points differ from image data with a uniform Cartesian distribution. Directly using convolution would result in the loss of a large amount of original prior feature information. Therefore, the neural network model employs a layered operation. The layered operation used in conjunction with the fully connected neural network FC1 is as follows:

[0086] S410, future total temperature T0, total pressure P0, Mach number Ma0, hydrogen injection position k, and injection flow rate L are input data into a fully connected neural network FC1 for nonlinear mapping.

[0087] S411. The nonlinear mapping result of the fully connected neural network FC1 is used to extract features and transform dimensions to obtain local features. That is, in the first layered convolution module, the nonlinear mapping result is sequentially passed through the Conv1, Bn1, ReLU, Conv2 and ReLU1 layers to output local features. In this step, the nonlinearity of the model is enhanced by the ReLU1 activation function to better encode the input feature information.

[0088] S412. The nonlinear mapping result of the fully connected neural network FC1 is adapted and transformed by a dynamic fusion module. This step is mainly to prevent the loss of feature information caused by the convolution process. The dynamic fusion module is introduced to perform adaptive dynamic transformation on the shallow features before connection in order to better fuse low-frequency and high-frequency features.

[0089] S413. The outputs of S411 and S412 are fused to obtain the first encoded feature corresponding to the global feature information.

[0090] It should be noted that since the output is the entire physical field data in three-dimensional space, the data dimension is high, and using a large number of fully connected layers will result in an extremely large number of parameters in the neural network model. Therefore, in order to reduce the number of model parameters, a dimension transformation is performed after the output of the layered convolution, converting the data of dimension (1, n) into a rectangle of dimension (a, b), where a*b equals n.

[0091] Furthermore, the hierarchical operations that work in conjunction with the fully connected neural network FC1 are as follows:

[0092] S420. Input the sampling point coordinates x, y, z into the position encoding module and perform polar coordinate encoding on the coordinates x, y, z. During the encoding process, angle and length displacement are used to encode the relative positional relationship between each grid point and the origin.

[0093] S421. Input the polar coordinate encoding result into the fully connected neural network FC2 for nonlinear mapping;

[0094] S422, the nonlinear mapping results of the fully connected neural network FC2 are output in parallel to the convolution Conv3 and convolution Conv4 in the second hierarchical convolution module for feature concatenation;

[0095] The features concatenated by the feature fusion Cat layer after S423, Conv3 and Conv4 are passed through the Sigmoid function to obtain the second encoded feature.

[0096] The feature decoding module decodes the first and second encoded features obtained from the first and second layered convolutional modules and outputs a multiphysics field. The specific operation flow is as follows:

[0097] S430. After the first encoded feature is processed by Conv5, the output of Conv5 is grouped and then fed into Conv6 and Conv7 for decoding.

[0098] S431, the second encoded feature is decoded by sequentially passing through convolutional Conv8, convolutional Conv9, and upscale layers;

[0099] S432. Perform pixel-by-pixel addition fusion (i.e., pixel-level addition) on the outputs of Conv6, Conv7, and Upscale layer, and output the fused information through Conv10 to output the multiphysics field.

[0100] S5. Based on physical mechanism models and neural network models, a three-dimensional flow field intelligent solution model based on physical information constraints is constructed, thereby replacing the numerical simulation calculation of supersonic combustion flow.

[0101] Specifically, the intelligent solution model for the three-dimensional flow field uses a physical mechanism model to constrain the prediction results of the neural network model, such as... Figure 2 As shown, the specific content is as follows: The neural network model constructed in S4 is used to establish a functional relationship between input and output, providing the preconditions for automatic differentiation of the physical mechanism model established in S3, thus achieving an effective combination of the physical mechanism model and the neural network model. Due to the strong non-convexity of the physical mechanism model, the entire model is difficult to converge. Therefore, the physical model is divided into two stages according to the differences in combustion characteristic distribution and flow characteristic distribution. This prevents the model from failing to achieve accuracy balance in end-to-end prediction of velocity fields and combustion product fields with large differences in characteristic distribution. In this embodiment, the physical mechanism model is grouped to reduce the model learning difficulty and accelerate the model convergence speed. The two stages are as follows:

[0102] In the first stage, a neural network model is used to predict the velocity fields in the x, y, and z directions, the pressure field p, and the Mach number field Mach, and a physical model is used. PY 1. Constrain it, specifically as follows:

[0103]

[0104] In the second stage, a neural network model was used to analyze the total temperature. , static temperature The concentration fields Ci of each component are predicted, and a physical model is used.PY 2. Constrain it, specifically as follows:

[0105]

[0106] The loss function in the three-dimensional flow field intelligent solution model loss It is characterized by the following formula:

[0107] ;

[0108] In the above formula, For physical models PY A deviation of 1, and ; For physical models PY The deviation of 2, and ; Indicates data-driven error, and It is characterized by the following formula:

[0109] ;

[0110] in, This represents labeled data, where m represents the total number of samples. This indicates the fitting result. u , v , w They represent in x direction, y direction, z velocity field in the direction, Ci Represents the concentration field of each component. Indicates the total temperature. Indicates static temperature. Mach Indicates Mach number. p Represents the density field. p It represents a pressure field.

[0111] Error analysis methods for intelligent solution models of three-dimensional flow fields based on physical information constraints are mainly adopted, using relative error (MAE) and root mean square error (MSE) to evaluate model performance, so as to achieve high-precision prediction of three-dimensional supersonic combustion flow and provide important technical support for the continuous and stable operation of engines.

[0112] To verify the effectiveness of the proposed method, a comparative verification was performed on two-dimensional supersonic combustion flow field data, as shown in Table 1.

[0113] Table 1 shows the average test performance on the two-dimensional combustion flow test set.

[0114]

[0115] As can be seen from Table 1, the hierarchical constraint method significantly reduces both the relative error (MAE) and root mean square error (MSE) of the model compared to the conventional simultaneous constraint method, which is sufficient to demonstrate the effectiveness of the hierarchical constraint method constructed in this invention in practical applications.

[0116] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0117] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for intelligently solving supersonic combustion flow fields based on hierarchical physical constraints, characterized in that, include: S1. Compare the experimental data obtained from the ground pulse combustion wind tunnel test with the simulation data obtained from the numerical simulation of supersonic combustion flow, and adjust the turbulence model parameters based on the comparison results before re-performing the numerical simulation until the simulation data matches the experimental data. S2. Based on the final turbulence model parameters obtained in S1, perform numerical simulation calculations of supersonic combustion flow under different incoming flow conditions to construct a three-dimensional combustion flow field dataset. S3. Based on the generation principle of combustion flow in scramjet engines, construct a physical mechanism model; S4. Based on the three-dimensional combustion flow field dataset, construct a neural network model based on hierarchical encoding and fusion decoding; S5. A physical mechanism model is used to constrain the prediction results of each layer of the neural network model in stages, so as to obtain a three-dimensional flow field intelligent solution model based on hierarchical physical information constraints. S6. Apply the three-dimensional flow field intelligent solution model to the prediction of three-dimensional supersonic combustion flow; In S3, the physical mechanism model includes: physical model PY 1 and physical model PY 2; Among them, physical model PY 1 is represented as: Physical Model PY 2 is represented as: In the above formula, ρ Represents the density field. u , v , w They represent in x direction, y direction, z velocity field in the direction, p Represents a pressure field. express x Shear stress in a plane express y Shear stress in a plane express z Shear stress in a plane express x plane and y Shear stress between planes express x plane and z Shear stress between planes express y plane and z Shear stress between planes Indicates the incoming Reynolds number, com Indicates the number of components. Ci Represents the concentration field of each component. i The first fuel i Number of components Indicates total temperature. Indicates static temperature. r Indicates specific heat ratio. Mach Indicates Mach number field; In S4, the fitting process of the neural network model is characterized by the following formula: In the formula, This indicates the fitting result. Represents a neural network model. x , y , z The values ​​represent the sampling point coordinates, t represents the non-stationary time, T0 represents the total incoming flow temperature, P0 represents the total pressure, Ma0 represents the Mach number, K represents the hydrogen injection location, and L represents the injection flow rate. u , v , w They represent in x direction, y direction, z velocity field in the direction, Ci Represents the concentration field of each component. Indicates total temperature. Indicates static temperature. Mach Indicates Mach field.

2. The intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints as described in claim 1, characterized in that, In S1, the method for adjusting the turbulence model parameters is as follows: S11. Based on the ground pulse combustion wind tunnel test, pressure sensors are used to obtain test data of combustion chamber wall pressure. S12. Based on the given parameters of the scramjet engine, use numerical simulation software to obtain simulation data of the entire three-dimensional space of the scramjet engine. S13. Based on the arrangement of pressure sensors in the ground pulse combustion wind tunnel test, select simulated wall pressure data from the simulation data at the same location as the scramjet engine. S14. Judge the consistency between the wall pressure test data and the wall pressure simulation data. If the judgment result is inconsistent, adjust the turbulence model parameters and return to S12. Otherwise, use the current turbulence model parameters as the final turbulence model parameters.

3. The intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints as described in claim 1, characterized in that, In S2, the three-dimensional combustion flow field dataset is constructed as follows: Three-dimensional numerical simulation calculations of scramjet engine combustion were performed based on the final turbulence model parameters. According to the wave system structure and flame distribution law of the multi-physics combustion field of scramjet engine, the sampling points in the three-dimensional flow field were obtained by adopting the segmented random sampling and local dense sampling strategies. The x, y, z coordinates of the sampling points, non-stationary time, total incoming temperature, total pressure, Mach number, hydrogen injection location, and injection flow rate are used as input data in the three-dimensional combustion flow field dataset. The velocity field, pressure field, Mach number field, total temperature, static temperature, and concentration field of each component in the x, y, and z directions in three-dimensional space are used as output data in the three-dimensional combustion flow field dataset.

4. The intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints as described in claim 3, characterized in that, The segmented random sampling and local dense sampling strategy refers to dividing the entire three-dimensional physical field into multiple sub-regions, determining the wave system structure and combustion complexity in the corresponding sub-regions based on the density gradient changes in each sub-region, and setting different numbers of sampling points according to the complexity of the corresponding sub-regions.

5. The intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints as described in claim 1, characterized in that, In S4, the neural network model includes: Parallel configuration of fully connected neural networks FC1 and FC2; The first and second layered convolutional modules are respectively set on the output side of the fully connected neural network FC1 and the fully connected neural network FC2; A feature decoding module that fuses the outputs of the first and second layered convolutional modules; Among them, on the output side of the fully connected neural network FC1, there is also a dynamic fusion module that is set in parallel with the first layered convolution module to adaptively and dynamically transform the nonlinear mapping of the fully connected neural network FC1. The nonlinear mapping data of the fully connected neural network FC1 comes from the inflow total temperature T0, total pressure P0, Mach number Ma0, hydrogen injection position k, injection flow rate L and unsteady time t in the three-dimensional combustion flow field dataset. At the front end of the fully connected neural network FC2, there is also a position encoding module that performs polar coordinate encoding of the coordinates x, y, z of the sampling points in the three-dimensional combustion flow field dataset.

6. The intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints as described in claim 5, characterized in that, The feature decoding module includes: Conv5 is a convolutional layer that performs convolution processing on the fused output of the first layer convolutional module and the dynamic fusion module; Convolutional layers Conv6 and Conv7 that group the output of convolutional layer Conv5; The output of the second layered convolutional module is sequentially processed by convolutional layers Conv8 and Conv9, and then by an upscale layer that performs upsampling on the output of Conv9. After pixel-wise addition and fusion processing of the outputs of convolutional layers Conv6, Conv7, and Upscale, the fused feature information is output by convolutional layer Conv10 using multi-physics field.

7. The intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints as described in claim 1, characterized in that, In S5, the phased constraint refers to: In the first stage, a neural network model was used to analyze the velocity and pressure fields in the x, y, and z directions. p The Mach number field is used for prediction, and the physical model in the physical mechanism model is used. PY 1. Constrain the prediction results; In the second stage, a neural network model was used to analyze the total temperature. , static temperature and the concentration field of each component Ci To make predictions, a physical model from the physical mechanism model is used. PY 2. Constrain the prediction results.

8. The intelligent solution method for supersonic combustion flow field based on hierarchical physical constraints as described in claim 7, characterized in that, In S5, the loss function in the three-dimensional flow field intelligent solution model loss It is characterized by the following formula: ； In the above formula, For physical models PY A deviation of 1, and ; For physical models PY The deviation of 2, and ; Indicates data-driven error, and It is characterized by the following formula: ； in, This represents labeled data, where m represents the total number of samples. This indicates the fitting result. u , v , w They represent in x direction, y direction, z velocity field in the direction, Ci Represents the concentration field of each component. Indicates total temperature. Indicates static temperature. Mach Indicates Mach number. ρ Represents the density field. p It represents a pressure field.

Images