Multi-engine vacuum plume field rapid calculation method based on deep learning

Abstract

The invention discloses a multi-engine vacuum plume field rapid calculation method based on deep learning, and relates to the technical field of aerospace propulsion systems and computational fluid mechanics, and the method comprises the following steps: sample data acquisition, data preprocessing, feature extraction, mapping relation construction, and flow field rapid calculation. According to the method, multi-dimensional feature representation is achieved through a symbol distance function, an identifier matrix and a free initial field, a deep learning agent model is constructed by combining data driving and physical constraint loss function design, the multi-engine plume complex interference effect is effectively processed, the prediction error of a key area is reduced, and the extreme working condition generalization ability is enhanced; meanwhile, traditional high-cost simulation is replaced, the flow field calculation efficiency is improved, and an efficient and accurate scheme is provided for spacecraft plume analysis and engineering application.

Description

Technical Field

[0001] The present invention relates to the technical field of aerospace propulsion systems and computational fluid dynamics, and specifically to a method for fast calculation of multi-engine vacuum plume fields based on deep learning. Background Art

[0002] 1. Vacuum plumes are high-temperature, high-speed flows of gas and particles ejected by spacecraft engines in vacuum environments such as low-Earth orbit or deep space. Their flow fields exhibit complex nonlinear dynamic characteristics. Obtaining vacuum plume fields, especially those under the combined action of multiple engines, often faces challenges such as the coupling of multi-scale turbulence and rarefied gas effects. Traditional methods rely on experimental observations or Direct Simulation Monte Carlo (DSMC) numerical simulations. However, these methods are limited by the large molecular free path span in a vacuum environment and the difficulty in modeling energy exchange. This makes it difficult to efficiently capture the non-equilibrium flow characteristics of the flow field, and full flow field reconstruction suffers from a conflict between temporal and spatial resolution and computational cost.

[0003] With the widespread application of artificial intelligence, fast flow field computation solutions based on deep learning—data-driven modeling using large amounts of simulation data obtained through traditional methods—have become a mainstream approach to flow field reconstruction. Currently, this data-driven approach uses methods such as normalization and grid interpolation to process spatial domain data. This approach relies on prior knowledge to establish a mapping between computational domain characteristics and the flow field, learns the spatial characteristics of the flow field, and generates a fast computational proxy model. This approach can rapidly compute multi-engine vacuum plume fields with complex nonlinear dynamics and provide initial conditions for evaluating plume thermodynamic effects.

[0004] 2. Currently, this method has been widely used in fluid mechanics, but its application in vacuum plume prediction is still relatively rudimentary. In particular, its application in multi-engine plume prediction faces the following challenges: First, multi-engine plume fields exhibit stronger nonlinear dynamic characteristics, with greater differences in the Knudsen magnitude of the flow and greater flow field complexity. The interaction of multiple nozzle plumes triggers dynamic interference effects such as shock wave collisions and turbulent mixing. These nonlinear superposition characteristics invalidate the traditional linear superposition assumption, necessitating the construction of a high-dimensional coupled dynamic model. Second, the characterization of multi-engine thrust and layout remains to be explored. Under multi-engine operating conditions, flow field characteristics (such as pressure gradient and vorticity) vary with nozzle layout and thrust ratio in a high-dimensional nonlinear manner. The stability of traditional input feature representations needs to be evaluated, and new thrust and layout feature representation methods are needed. Finally, existing vacuum plume reconstruction models lack physical constraints. Purely data-driven models are easily limited by the distribution of training data, resulting in a sharp drop in prediction reliability under extreme vacuum plume conditions. Purely physically constrained models, however, struggle to capture complex boundary conditions due to the extremely complex differential and integral terms in the vacuum plume governing equations or the inherent errors in simplified equations.

[0005] In view of this, this patent proposes a fast calculation method for multi-engine vacuum plume fields based on deep learning, which aims to use data-driven methods combined with partial control equation constraints to extract the initial free plume field that conforms to the engine layout as one of the input features, so as to achieve fast and accurate calculation of multi-engine vacuum plume fields. The fast calculation method for multi-engine vacuum plume fields based on deep learning pre-processes experimental or simulated data in the spatial domain, constructs a mapping relationship between engine layout and thrust characteristics, boundary condition characteristics and data, and realizes the calculation of multi-engine vacuum plume fields with different layouts and thrusts, which provides the basis for vacuum plume flow characteristic analysis and mechanical and thermal effect evaluation. Summary of the Invention

[0006] The purpose of the present invention is to provide a fast calculation method for multi-engine vacuum plume fields based on deep learning to solve the problems raised in the above background technology.

[0007] To solve the above technical problems, the present invention provides a method for fast calculation of multi-engine vacuum plume fields based on deep learning, comprising the following steps:

[0008] Sample data acquisition: Based on experiments or DSMC simulations, the flow field conservation information of the engine vacuum plume field for a given altitude, boundary conditions, engine layout, and thrust is obtained, including the density field, momentum density field, and energy density field. At the same time, the plume field in the unrestricted free state corresponding to each engine thrust is obtained as the free initial field, forming a data set that includes the number of operating conditions, the number of conservation quantities, and the number of spatial grids.

[0009] Data preprocessing: normalizing or standardizing the flow field conservation information so that the conservation distribution is concentrated between (0, 1);

[0010] Feature extraction: abstract the spatial topology and engine layout information into a signed distance function (SDF), abstract the boundary condition information into an identifier matrix, and abstract the engine thrust information into a normalized free initial field;

[0011] Mapping relationship construction: A deep learning proxy model is trained using data-driven modeling constraints, multi-scale characteristic constraints, and continuity equation constraints to establish a mapping relationship between input features and flow field data. The loss function includes a traditional data-driven modeling term, a multi-scale characteristic constraint term with the gradient length Knudsen number as a weighting factor, and a continuity equation loss term.

[0012] Fast flow field calculation: Load the trained deep learning agent model and input the characteristics under any given altitude, engine layout and thrust conditions for fast calculation of the vacuum plume field.

[0013] Furthermore, the signed distance function SDF is defined as:

[0014]

[0015] in, is the boundary of the spatial domain Ω, Ω c represents the complement of Ω, From the spatial point x to the boundary The minimum distance; and the SDF is normalized and its absolute value is less than 1. ,

[0016] Furthermore, the identifier matrix is ​​assigned different numbers according to different types of boundaries in the computational domain and different regions of the flow field according to spatial positional relationships, and different numbers represent different types of boundaries or plume flow regions.

[0017] Furthermore, the parameters of the free initial field are normalized according to the data preprocessing method, and its structural characteristics and parameter distribution characteristics are used to represent the thrust information of the engine. If there are m engines and n different thrusts, there are a total of m×n free initial fields in the input.

[0018] Furthermore, the method for processing the conserved quantity in data preprocessing is:

[0019] Density field normalization:

[0020] Momentum density field normalized:

[0021] Energy density field normalization: Among them, the upper wavy line represents the normalized conserved quantity, ∈ and δ are small quantities with the dimensions of density and energy density, respectively. and are the mean and standard deviation of the outlet parameters of the engine at corresponding thrust.

[0022] Furthermore, the gradient length Knudsen number Kn is calculated as:

[0023]

[0024] Where λ is the molecular mean free path, q is the flow field parameter and is set to density, is the second norm of the gradient of the flow field parameters.

[0025] Furthermore, the loss function expression is:

[0026]

[0027] Where F is an arbitrary flow field parameter, the superscripts P and T represent the predicted value and the true value, respectively, and N is the total number of data points in the sample space.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. Efficient and stable feature representation supports complex flow field modeling:

[0030] 1. Multi-dimensional feature abstraction technology: This technology converts spatial topology and engine layout information into standardized numerical features, accurately describing the computational domain boundaries and the relative positions of the engines. This technology provides a stable spatial structure input for the model and resolves the issue of inconsistent layout feature representation in traditional methods.

[0031] The identifier matrix digitally encodes the boundary conditions and flow field areas, clearly distinguishing different areas such as open boundaries, solid boundaries, and core flow areas, enabling the model to adapt to the flow field characteristics under different boundary conditions and improve generalization capabilities.

[0032] The free initial field characterizes the engine thrust information and stably transmits the dynamic characteristics of the thrust through the structural characteristics of the unrestricted free plume field (such as the position and angle of the compression wave) and the parameter distribution characteristics, avoiding the instability of traditional thrust representation methods under high-dimensional nonlinear conditions and providing reliable input for the modeling of multi-engine coupling effects.

[0033] 2. Feature normalization and stability improvement: Targeted normalization processing is performed on the density field, momentum density field, energy density field, and free initial field parameters. Through logarithmic transformation and standardization methods, the data distribution is concentrated in the (0,1) interval, significantly improving the learning efficiency of the deep learning model and its adaptability to extreme working conditions.

[0034] 2. Multi-constraint training strategy to enhance the physical reliability and accuracy of the model:

[0035] 1. Fusion of data-driven and physical constraints: The loss function introduces the continuity equation loss term and directly embeds the basic conservation laws of fluid mechanics. This makes up for the lack of physical constraints in pure data-driven models, ensures that the predicted flow field meets the mass conservation conditions, and improves the prediction reliability under extreme working conditions.

[0036] The gradient length Knudsen number (Kn) weighted constraint term targets the coupling problem of multi-scale turbulence and rarefied gas effects. By taking the ratio of the second norm of the density gradient to the molecular mean free path, it enhances the model's ability to capture multi-scale features of key areas such as shock waves and expansion waves, significantly reduces the prediction error of compression waves and shock wave areas, and improves the reconstruction accuracy of non-equilibrium characteristics of the flow field.

[0037] 2. High-dimensional coupled dynamics modeling capabilities: Combining the ConvNeXt block convolutional neural network with the U-Net structure, a nonlinear mapping relationship is constructed from high-dimensional input to flow field data. This effectively handles dynamic interference effects such as shock wave collisions and turbulent mixing caused by the interaction of multiple engine plumes, breaking through the limitations of traditional linear superposition assumptions and achieving efficient modeling of complex flow fields.

[0038] 3. Rapid calculation and engineering application value:

[0039] 1. Computational efficiency and cost optimization: By replacing traditional DSMC numerical simulations with deep learning proxy models, the computational cost of multi-engine vacuum plume fields is significantly reduced, the flow field reconstruction time is shortened, and the real-time analysis requirements for spacecraft plume thermodynamic effect assessment and orbit control are met, providing an efficient tool for engineering applications.

[0040] 2. Generalization and adaptability to extreme operating conditions: The training method that integrates physical constraints reduces the model's dependence on the distribution of training data and improves its generalization ability in vacuum environments with large Knudsen number spans and complex flow states. This method is particularly suitable for extreme operating conditions such as deep space exploration and multi-engine collaboration in low-Earth orbit, providing reliable flow field data support for spacecraft design and mission planning.

[0041] 4. Industry promotion:

[0042] 1. Methodological breakthrough: This paper proposes an interdisciplinary approach that combines data-driven and physical constraints, providing a new modeling approach for the vacuum plume field. This approach can be extended to other complex flow field calculation scenarios, promoting the deep integration of fluid mechanics and artificial intelligence.

[0043] 2. Basis for real-time analysis and optimization: The established rapid calculation model lays the foundation for real-time analysis of multi-engine plume field structure and thermodynamic effects, assists in the development of key technologies such as spacecraft propulsion system optimization design and plume interference avoidance, and enhances the autonomy and reliability of my country's space missions.

[0044] In summary, the present invention achieves fast, accurate and reliable calculation of multi-engine vacuum plume fields through new feature representation, multi-constraint training and efficient model architecture, effectively solving the bottleneck problem of existing technologies and having significant scientific significance and engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of the module composition of a method for fast calculation of multi-engine vacuum plume fields based on deep learning in the present invention;

[0046] Figure 2 This is a schematic diagram of the main steps of a method for fast calculation of multi-engine vacuum plume fields based on deep learning in the present invention;

[0047] Figure 3 This is a schematic diagram of a method for representing spatial topology and engine layout characteristics in a deep learning-based method for rapidly calculating a multi-engine vacuum plume field;

[0048] Figure 4 This is a schematic diagram of the three-dimensional calculation domain of a method for fast calculation of multi-engine vacuum plume fields based on deep learning in the present invention;

[0049] Figure 5 This is a schematic diagram of an engine thrust characteristic representation method for a fast calculation method of a multi-engine vacuum plume field based on deep learning in the present invention.

[0050] In the figure: 101, sample data acquisition module; 102, data preprocessing module; 103, feature extraction module; 104, mapping relationship construction module; 105, flow field fast calculation module. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] See also Figure 1-Figure 5 , the present invention provides a technical solution:

[0053] See Figure 1-Figure 5 As shown in Figure 1, a fast calculation method for multi-engine vacuum plume fields based on deep learning is:

[0054] The ConvNeXt block convolutional neural network based on data-driven means and U-Net structure is used to quickly reconstruct a given altitude, engine layout and thrust, and includes a sample data acquisition module 101, a data preprocessing module 102, a feature extraction module 103, a mapping relationship construction module 104, and a flow field fast calculation module 105.

[0055] The sample data acquisition module 101 is primarily used to generate sample flow field data. Data generation can be based on experiments or DSMC simulations. The engine vacuum plume field for a given altitude, boundary conditions, engine layout, and thrust is simulated to obtain information on conserved quantities in the flow field, including the density field ρ, a conserved quantity in the continuity equation; the momentum density field ρU, a conserved quantity in the momentum equation; and the energy density field ε, a conserved quantity in the energy equation. In addition to obtaining flow field data for sample operating conditions, for each engine thrust, the plume field in an unrestricted free state, known as the free initial field, is also required for subsequent thrust characterization. Taking the three-dimensional case as an example, the resulting data set is structured as N×5×x×y×z, where N is the number of operating conditions in the data set, 5 is the number of predicted conserved quantities for the three-dimensional case, and x, y, and z are the number of spatial grids in the three directions of the Cartesian coordinate system.

[0056] The data preprocessing module 102 is mainly used to normalize or standardize the spatial domain data obtained by the sample data acquisition module 101 to obtain a normalized conservation quantity whose distribution is concentrated between (0, 1). The preprocessing method for the conservation quantity is:

[0057]

[0058]

[0059] Among them, the upper wavy line represents the normalized conserved quantity, ∈ and δ are small quantities with the dimensions of density and energy density, respectively. and are the mean and standard deviation of the outlet parameters of the engine at corresponding thrust.

[0060] The feature extraction module 103 mainly abstracts the spatial topology information, boundary condition information, engine layout and thrust information of each working condition into input features with a certain degree of stability. First, the spatial topology information and engine layout information are abstracted into a signed distance function (SDF):

[0061]

[0062] in, is the boundary of the spatial domain Ω, and Ω c represents the complement of Ω, for any x∈X,

[0063]

[0064] Among them, inf represents the lower bound, the metric space X is the entire computational domain, and its subset Ω represents the flow field excluding the engine. The SDF is also normalized, and its absolute value is less than 1. The boundary condition information is abstracted as an identifier matrix, in which different types of boundaries of the computational domain and different regions of the flow field are assigned different numbers according to the spatial position relationship. Different numbers represent different types of boundaries or plume flow areas. The engine thrust information is abstracted as the free initial field calculated in module 101. The parameters in the initial field are also normalized according to the method of module 102. The structural characteristics of the free initial field, such as the position and angle of the compression wave, and the parameter distribution characteristics can more stably represent the thrust information of the engine in a given working condition. Taking a certain working condition as an example, there are three engines in this working condition, one is located in the center of the computational domain, and the other two are placed side by side with it in the x and y directions, and are separated from it by an outlet diameter in the z direction. The right side of the computational domain is a solid boundary, and the rest are open boundaries. The SDF section generated by the above method for this working condition is as follows Figure 3 As shown on the left, the identifier matrix section divided by the engine, open boundary, solid boundary, core flow area, plume strong interaction area, large angle flow area and recirculation area is as follows Figure 3 As shown on the right, the computational domain and the position of the above SDF section are as follows Figure 4 The free initial section of the engine under a certain thrust is shown as Figure 5 As shown in the figure, the initial field does not contain the engine spatial position and layout information. If there are m engines and n different thrusts, there are a total of m×n free initial fields in the input.

[0065] The mapping relationship building module 104 mainly trains a deep learning agent model that can establish a mapping relationship between input features and flow field data. During the training process, the loss function is set as follows

[0066]

[0067] Where F is an arbitrary flow field parameter, the superscripts P and T represent the predicted value and the true value respectively, N is the total number of data points in the sample space, and Kn is the gradient length Knudsen number, which is calculated as follows

[0068]

[0069] Wherein, λ is the molecular mean free path, q is an arbitrary flow field parameter, which is set to density in the embodiment of the present invention. is the second norm of the flow parameter gradient. The specific meanings of the various terms in the loss function are as follows: the first term is the traditional data-driven modeling term; the second term adds the gradient length Knudsen number as a weighting factor, which characterizes the gas compressibility effect. For example, in shock or expansion wave regions, where the density gradient is significant, local minima at characteristic scales can identify shock wave thickness or rarefied transition regions, enhancing the model's ability to analyze the multi-scale characteristics of the plume field; the third term is the continuity equation loss term, which is used to provide continuity physical constraints.

[0070] The flow field fast calculation module 105 mainly realizes the fast calculation of the vacuum plume field under any given altitude, engine layout and thrust conditions by loading the trained deep learning agent model, wherein each input feature is obtained through the feature extraction module 103.

[0071] Based on the five modules of sample data acquisition module 101, data preprocessing module 102, feature extraction module 103, mapping relationship construction module 104, and flow field rapid calculation module 105, the rapid calculation of multi-engine vacuum plume fields can be carried out. The general workflow is as follows:

[0072] Step 201: For the selected sample operating condition, the spatial distribution data of the vacuum plume field is obtained by means of experiments or numerical simulations. The parameters include the conservation quantities of the continuity equation, momentum equation, and energy equation. At the same time, the free initial field of the thrust of each engine in the sample operating condition is calculated to obtain the spatial domain distribution of the sample data.

[0073] In step 202 , the spatial domain data generated by the sample data acquisition module 101 is imported into the data preprocessing module 102 for normalization or standardization to improve the learning ability and generalization of the deep learning model.

[0074] Step 203 , extracting the height, boundary conditions, engine layout and thrust characteristics of the sample operating condition, obtaining the signed distance function, identifier matrix and normalized free initial field, and importing the characteristic data and the normalized flow field data in step 202 into the mapping relationship construction module 104 .

[0075] Step 204: Construct a mapping relationship between the features and the plume field data. The data generated in steps 202 and 203 are used to train a deep learning agent model using data-driven modeling constraints, multi-scale characteristic constraints, and continuity equation constraints to obtain a mapping relationship between the altitude, engine layout and thrust features and the plume field data.

[0076] Step 205 , a fast calculation model of the plume field is obtained, and features are established using the feature extraction module 103 and input into the proxy model trained by the mapping relationship construction module 104 to realize the calculation from the features of any working condition to the flow field.

[0077] Summarize:

[0078] This paper proposes a fast calculation method for multi-engine vacuum plume fields based on deep learning, which effectively solves the challenges of existing technologies in multi-engine plume field prediction:

[0079] Model Construction and Training Methodology: A deep learning agent model is trained using a data-driven, multi-scale feature-constrained, and continuity-equation-constrained approach. A modular architecture encompasses sample data acquisition, preprocessing, feature extraction, mapping relationship construction, and rapid flow field calculation. The loss function incorporates traditional data-driven modeling, a multi-scale feature-constrained term weighted by the gradient length Knudsen number, and a continuity-equation loss term. This allows for rapid computation of complex vacuum plume fields across multiple engines, enhancing the model's ability to analyze the multi-scale characteristics of the plume field while providing continuous physical constraints.

[0080] Thrust feature representation: The free initial field is used to characterize the engine thrust information. The free initial field is the plume field in the unrestricted free state of the engine. After normalization, its structural characteristics (such as the position and angle of the compression wave) and parameter distribution characteristics can stably provide thrust characteristics that can be distinguished by the deep learning agent model.

[0081] Feature Representation Method: A method based on the signed distance function (SDF), an identifier matrix, and the engine free initial field (FFI) is proposed to represent altitude, engine layout, and thrust information. The SDF abstracts spatial topology and engine layout information, the identifier matrix abstracts boundary conditions, and the FFI abstracts engine thrust information. This enables the training of a deep learning agent model for the rapid computation of multi-engine vacuum plume fields, providing a foundation for real-time analysis of complex multi-engine plume structures and their thermodynamic and thermal effects.

[0082] Error reduction method: A multi-scale feature constraint and continuity equation constraint training method is proposed to reduce the prediction error of the key compression wave and shock wave regions. The gradient length Knudsen number in the multi-scale feature constraint can be directly calculated from the sample data. This factor characterizes the gas compressibility effect and can identify the shock wave thickness or rarefied transition zone, thereby enhancing the model's ability to analyze the multi-scale characteristics of the plume field.

Claims

1. A fast calculation method for multi-engine vacuum plume field based on deep learning, characterized by: The following steps are involved: Sample data acquisition: Based on experiments or DSMC simulations, the flow field conservation information of the engine vacuum plume field for a given altitude, boundary conditions, engine layout, and thrust is obtained, including the density field, momentum density field, and energy density field. At the same time, the plume field in the unrestricted free state corresponding to each engine thrust is obtained as the free initial field, forming a data set that includes the number of operating conditions, the number of conservation quantities, and the number of spatial grids. Data preprocessing: normalize or standardize the flow field conservation information so that the conservation distribution is concentrated between (0, 1); Feature extraction: abstract the spatial topology and engine layout information into a signed distance function (SDF), abstract the boundary condition information into an identifier matrix, and abstract the engine thrust information into a normalized free initial field; Mapping relationship construction: The deep learning agent model is trained using data-driven modeling constraints, multi-scale characteristic constraints, and continuity equation constraints to establish a mapping relationship between input features and flow field data. The loss function includes traditional data-driven modeling terms, multi-scale characteristic constraint terms with the gradient length Knudsen number as the weight factor, and continuity equation loss terms. Fast flow field calculation: Load the trained deep learning agent model and input the characteristics under any given altitude, engine layout and thrust conditions for fast calculation of the vacuum plume field.

2. The method for rapid calculation of multi-engine vacuum plume fields based on deep learning according to claim 1, characterized in that: The signed distance function SDF is defined as: in, is the boundary of the spatial domain Ω, Ω c represents the complement of Ω, From the spatial point x to the boundary The minimum distance; and the SDF is normalized and its absolute value is less than 1.

3. The method for rapid calculation of multi-engine vacuum plume fields based on deep learning according to claim 1, characterized in that: The identifier matrix assigns different numbers according to the spatial position relationship of different types of boundaries in the computational domain and different areas of the flow field. Different numbers represent different types of boundaries or plume flow areas.

4. The method for rapid calculation of multi-engine vacuum plume fields based on deep learning according to claim 1, characterized in that: The parameters of the free initial field are normalized according to the data preprocessing method, and its structural characteristics and parameter distribution characteristics are used to represent the thrust information of the engine. If there are m engines and n different thrusts, there are a total of m×n free initial fields in the input.

5. The method for rapid calculation of multi-engine vacuum plume fields based on deep learning according to claim 1, characterized in that: The method for processing conserved quantities in data preprocessing is: Density field normalization: Momentum density field normalized: Energy density field normalization: Among them, the upper wavy line represents the normalized conserved quantity, ∈ and δ are small quantities with the dimensions of density and energy density, respectively. and are the mean and standard deviation of the outlet parameters of the engine at corresponding thrust.

6. The method for rapid calculation of multi-engine vacuum plume fields based on deep learning according to claim 1, characterized in that: The gradient length Knudsen number Kn is calculated as: Where λ is the molecular mean free path, q is the flow field parameter and is set to density, is the second norm of the gradient of the flow field parameters.

7. The method for rapid calculation of multi-engine vacuum plume fields based on deep learning according to claim 1, characterized in that: The loss function expression is: Where F is an arbitrary flow field parameter, the superscripts P and T represent the predicted value and the true value, respectively, and N is the total number of data points in the sample space.

Images