Efficient cross-flow-state aerodynamic force heat effect prediction and database construction method based on data-driven deep learning
Through high-precision CFD simulation and experimental data collection, the establishment of a local coordinate system and dimensionless processing, the construction of a dynamic database of cross-flow state data, and the use of deep learning models to train the mapping relationship between aerodynamic forces and thermal effects, the problems of insufficient data utilization and insufficient prediction accuracy in existing technologies are solved, and efficient and accurate prediction of aerodynamic and thermal effects is achieved.
Patent Information
- Application Number
- CN202510511855.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-23
- Publication Date
- 2025-09-09
AI Technical Summary
Existing deep learning-based aerodynamic thermal effect prediction technology has problems such as insufficient data utilization, poor generalization ability, imprecise local feature extraction, and imperfect dynamic data updating, resulting in waste of computing resources and insufficient prediction accuracy.
Through high-precision CFD simulation and experimental data acquisition, a universal local coordinate system is established and dimensionless processing is performed, a dynamic database of cross-flow state data is constructed, a deep learning model is used to train the mapping relationship between local aerodynamic forces and thermal effects, and the CFD calculation is corrected through a feedback mechanism.
It achieves efficient and accurate prediction of local aerodynamic and thermal effects, reduces computing resource consumption, improves the adaptability and robustness of the model in cross-flow scenarios, and supports rapid engineering response.
Smart Images
Figure CN120611652A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of aerodynamic thermal effect prediction, and specifically to an efficient cross-flow state aerodynamic thermal effect prediction and database construction method based on data-driven deep learning. Background Art
[0002] 1. Brief description of the prior art:
[0003] For the past few decades, the prediction of aerodynamic and thermal effects has primarily relied on computational fluid dynamics (CFD) numerical simulations and experimental testing. In recent years, with the development of artificial intelligence and deep learning technologies, some research efforts have begun to explore data-driven approaches to accelerate and optimize traditional CFD calculations. Traditional CFD methods solve the Navier–Stokes equations and construct discrete computational models using finite volume, finite difference, or finite element methods to address aerodynamic forces, thermal effects, and boundary layer behavior in complex flow fields. These methods are representative in the following areas:
[0004] High-precision prediction: When the physical laws of the flow field and boundary conditions are clear, CFD can accurately simulate the interaction between fluids and solids in plume contamination, hypersonic reentry, and other complex conditions.
[0005] Local details: Through meticulous meshing and advanced solution methods, CFD can reveal key information such as flow field structure, pressure distribution, and temperature gradient in local areas.
[0006] However, such methods generally have defects such as long calculation time, high requirements for computing resources, and complex mesh generation and solution processes, which are not conducive to the needs of rapid response or prediction in a large range of parameter spaces for multiple flow states in engineering.
[0007] In recent years, deep learning-based technologies have gradually emerged in the field of flow field prediction. Some research efforts have attempted to use deep neural networks as proxy models, directly predicting flow field distribution or local aerodynamic and thermal effect data by inputting information such as geometric parameters and boundary conditions. The basic ideas of this type of method include:
[0008] Data-driven modeling: Using large-scale CFD simulation data and some experimental data, a deep neural network model is constructed through supervised learning to map geometric or boundary information to the local flow field structure or aerodynamic thermal effect distribution;
[0009] Model training and generalization: Offline training is performed on existing datasets to form a preliminary prediction model, which is then fine-tuned using a small amount of new data to adapt to specific flow conditions.
[0010] Multi-scale processing: Some methods attempt to capture detailed features of different regions (such as the boundary layer, wake, and shock wave regions) through multi-scale network structures, but most work is limited to overall flow field prediction.
[0011] Overall, existing deep learning-based aerodynamic thermal effect prediction methods primarily rely on geometry, boundary conditions, and global flow field parameters as input, directly inferring the aerodynamic thermal effects of the entire flow field or key local regions. While these technologies have, to a certain extent, broken through the computational bottlenecks of traditional CFD and achieved near-real-time prediction, they also expose the following shortcomings:
[0012] 2. Objective shortcomings of existing technologies:
[0013] 2.1. Insufficient data utilization:
[0014] Existing technologies mostly focus on using geometric or boundary condition data as network input, ignoring the large amount of high-fidelity flow field data accumulated during CFD calculations and experiments. This results in the inadequate exploitation of valuable local flow field information, characteristic data within the local coordinate system, and dimensionless processing information, which in turn limits the accuracy and robustness of the model in predicting local details.
[0015] 2.2. Poor generalization ability leads to waste of computing resources:
[0016] Current deep learning models are typically trained for specific flow conditions. To adapt to diverse flow environments (such as hypersonic reentry, plume contamination, or other extreme conditions), a usable model must be constructed using extensive new CFD simulations to construct training datasets. This not only wastes computing resources but also makes the model's generalization capabilities heavily dependent on the coverage of the training data, making it difficult to adapt to rapidly changing operating conditions in real-world engineering.
[0017] 2.3. Insufficient fusion of numerical simulation and experimental data:
[0018] While some studies have attempted to use experimental data for network training, a unified processing framework compatible with both high-fidelity CFD and experimental data is generally lacking. Different data sources vary significantly in terms of dimensionality, accuracy, and acquisition methods. Directly using these as input for model training often introduces noise and inconsistencies, impacting the reliability of the final predictions. Existing methods often struggle to seamlessly integrate different data sources, resulting in insufficient adaptability of trained models to diverse data.
[0019] 2.4. Local feature extraction is not refined:
[0020] Existing methods often use a global input model, which cannot accurately capture the detailed changes in the local flow field even when considering local effects. This crude local feature extraction method often leads to large errors in the prediction of local aerodynamic thermal effects (such as shock wave boundaries and critical areas of thermal protection systems), thus affecting the accuracy of the overall prediction results.
[0021] 2.5. Dynamic updating of data sets and training of proxy models are not perfect:
[0022] In practical engineering applications, flow conditions are often in a constantly changing dynamic process. Existing methods struggle to improve models through online or periodic dataset updates, and they lack the ability to implement a feedback mechanism between surrogate models and CFD numerical simulations. Consequently, when encountering new flow conditions or extreme operating conditions, the models often fail to respond quickly and adjust predictions, impacting engineering decisions.
[0023] In summary, while existing deep learning-based aerodynamic and thermal effect prediction technologies have improved computational efficiency to a certain extent, they still suffer from objective shortcomings such as low data utilization, insufficient generalization capability, inaccurate local feature capture, and incomplete dynamic data updates. In view of this, an efficient cross-flow regime aerodynamic and thermal effect prediction and database construction method based on data-driven deep learning is proposed to overcome the above-mentioned problems. The aim is to construct an efficient prediction method based on data-driven deep learning. This method uses local information extraction, a unified local coordinate system, and dimensionless processing, and utilizes database management to implement surrogate model training. This method aims to significantly improve computational efficiency while ensuring high-precision predictions, and achieve effective utilization of cross-basin data and model generalization. Summary of the Invention
[0024] The purpose of the present invention is to provide an efficient cross-flow aerodynamic thermal effect prediction and database construction method based on data-driven deep learning to solve the problems raised in the above background technology.
[0025] To solve the above technical problems, the present invention provides a method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning, comprising the following steps:
[0026] Through high-precision CFD simulation and experiments, the aerodynamic and thermal effect data of each area in the flow field are collected;
[0027] For local areas under different flow regimes, a universal local coordinate system is established, and the scale differences between different data sources are eliminated through local dimensionless processing;
[0028] Calculate the local Knudsen number Kn of local flow field data, classify and store the data according to flow state, and build a dynamic database of cross-flow state data;
[0029] Based on data preprocessing and local feature extraction, a deep neural network is used to model the local aerodynamic and thermal effect mapping and train a deep learning agent model.
[0030] The trained deep learning model is used for real-time or quasi-real-time prediction of local aerodynamic and thermal effects. The prediction results are fed back to the CFD calculation through a feedback mechanism to correct the gas-solid interaction parameters.
[0031] Furthermore, a universal local coordinate system is established. Specifically, based on the assumption that the aerodynamic thermal effect is strongly correlated with the off-wall node, a simplified model is constructed to simplify the process of the flow field generating the aerodynamic thermal effect into the effect of the incoming flow on the bottom surface position of the microelement within a microelement, which is defined by the calculation node parameters contained in the microelement; a local coordinate system is established inside the microelement through the characteristic velocity V, which is the velocity component v along the shear stress direction of the calculation node near the inlet surface of the microelement i The weighted average of is calculated as:
[0032]
[0033] Among them, n i is the molecular number density or fluid density at the node.
[0034] Furthermore, a local dimensionless treatment is performed, specifically: the density ρ and temperature T near the infinitesimal inlet are used to dimensionlessly transform the flow parameters, and the formula is:
[0035]
[0036] In the formula, the superscript * represents a dimensionless parameter.
[0037] Furthermore, when storing data in the dynamic database of cross-flow regime data, the local Knudsen number Kn is used to set a label for the characteristics of the local flow. The calculation formula of the local Knudsen number Kn is:
[0038]
[0039] Where λ is the molecular mean free path, U is the flow field physical quantity, and the flow field velocity and density are used to divide the flow state.
[0040] Furthermore, when training the deep learning agent model, the cross-flow data dynamic database detects the flow characteristics Kn of the data set used, and filters and provides data with similar Kn values from the database to supplement the training data set to achieve data enhancement.
[0041] Furthermore, the specific process of data enhancement is as follows: establish a universal local coordinate system, perform local dimensionless processing, and complete the preprocessing of new data; use the preprocessed new data to construct a deep learning model training data set; store the new data in a dynamic database of cross-flow state data, and calculate the local flow field characteristics Kn; screen the local flow field data output with similar flow field characteristics in the database and add them to the deep learning model training data set; the deep learning agent model uses the enhanced training data set to complete training and learn the complex nonlinear mapping relationship between the local flow field and aerodynamic and aerothermal effects.
[0042] Furthermore, a deep learning agent model is constructed and trained to learn the complex nonlinear mapping relationship between fluid aerodynamics, aerodynamic thermal effects and local flow fields. There are no special requirements for the structure of the agent model used, as long as the input and output modes are met.
[0043] Furthermore, the CFD calculation is fed back through a feedback mechanism. Specifically, the prediction results of the deep learning model are used to correct the gas-solid coupling process in the CFD calculation, thereby improving the overall numerical simulation accuracy and efficiency.
[0044] Compared with the prior art, the present invention has the following beneficial effects:
[0045] 1. Efficient Proxy Model Construction: Through high-precision CFD simulation and experimental data collection, combined with deep learning proxy model training, efficient learning of the nonlinear mapping relationship between local aerodynamic forces and thermal effects is achieved. This avoids the complex mesh generation and long calculation time problems of traditional CFD, significantly improves prediction efficiency, and meets the needs of rapid engineering response.
[0046] Compared with existing technologies that rely on global input modes of geometry or boundary conditions, this method focuses on local regional characteristics and directly captures the complex relationship between aerodynamic thermal effects and local flow fields through data-driven modeling, providing a unified and efficient framework for cross-flow state prediction.
[0047] 2. Improve local prediction accuracy and robustness: A dedicated coordinate system is constructed for local areas with different flow states. The local coordinate system is established based on the characteristic velocity within the microelement, accurately mapping the strong correlation between aerodynamic and thermal effects and the off-wall nodes. This solves the problem of imprecise local feature extraction in existing technologies, enabling the model to capture detailed changes in key areas such as the boundary layer and shock waves, significantly improving the prediction accuracy of local aerodynamic and thermal effects.
[0048] The density and temperature near the microelement inlet are used to non-dimensionalize the flow parameters, eliminating the dimensional differences of multi-source data (CFD simulation and experimental data), forming a unified data scale, enhancing the model's adaptability to different flow states (such as hypersonic reentry and plume contamination), and avoiding noise and prediction errors caused by dimensional inconsistency.
[0049] 3. Addressing Data Scarcity and Improving Training Efficiency: Data is partitioned and labeled using the local Knudsen number (Kn). Historical data is selected based on flow state similarity and added to the training set for data augmentation. This avoids the waste of resources required to simulate numerous new CFD models under new conditions. This dynamic update mechanism effectively alleviates data scarcity and improves the generalization capabilities of the surrogate model under untrained flow states.
[0050] The amount of processed data stored is much smaller than that of traditional CFD results (such as tens of MB storing thousands of sets of data), and it focuses on the key parameters of gas-solid interaction, reducing data storage and processing costs while ensuring the high pertinence and effectiveness of model training data.
[0051] 4. Multi-source data fusion and cross-flow regime adaptability: The generalization of local coordinate systems and dimensionless processing supports the seamless integration of CFD simulation and experimental data, solving the problem of insufficient fusion of different data sources in existing technologies, improving the robustness of the model in cross-flow regime scenarios, and enabling it to adapt to diverse flow environments such as hypersonic speeds and extreme working conditions.
[0052] 5. Mutual Feedback Mechanism between the Proxy Model and CFD: Prediction results feed back into CFD calculations to correct gas-solid coupling parameters, forming a closed loop of "prediction-correction-optimization." This mechanism not only improves the prediction accuracy of the deep learning model but also enhances the overall efficiency and accuracy of numerical simulations by optimizing the gas-solid interaction process in CFD simulations, achieving a complementary advantage between data-driven methods and traditional CFD.
[0053] 6. Significantly reduce computing resource consumption: Through data enhancement, dynamic database updates, and efficient proxy model training, it avoids the high computing resource requirements of traditional CFD and the reliance of existing deep learning methods on large-scale new data sets. While ensuring high-precision predictions, it significantly reduces computing costs and meets the needs of rapid prediction of multiple flow states and a wide range of parameter spaces in engineering.
[0054] In summary, this method constructs a prediction system that is both efficient, accurate, and widely adaptable through deep mining of local information, multi-source data fusion, efficient data utilization, and mutual feedback mechanism. It effectively solves the problems of insufficient data utilization, poor generalization ability, and inaccurate capture of local details in existing technologies, and provides a new solution for the prediction of aerodynamic thermal effects under complex flow conditions. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 Schematic diagram of the module composition of an efficient cross-flow regime aerodynamic thermal effect prediction and database construction method based on data-driven deep learning of the present invention;
[0056] Figure 2Schematic diagram of the main steps of a method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning in the present invention;
[0057] Figure 3 Schematic diagram of the construction of a universal local aerodynamic thermal effect mapping coordinate system in an efficient cross-flow regime aerodynamic thermal effect prediction and database construction method based on data-driven deep learning of the present invention;
[0058] Figure 4 This is a schematic diagram of the main links of the construction and management of a dynamic database of cross-flow state data in an efficient cross-flow state aerodynamic thermal effect prediction and database construction method based on data-driven deep learning in the present invention. DETAILED DESCRIPTION
[0059] 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.
[0060] See also Figure 1-Figure 4 , the present invention provides a technical solution:
[0061] See Figure 1-Figure 4 As shown in the figure, an efficient cross-flow aerodynamic thermal effect prediction and database construction method based on data-driven deep learning is proposed:
[0062] It includes: a universal local aerodynamic thermal effect mapping coordinate system construction and local physical quantity dimensionless module 101, a deep learning agent model module 102, and a cross-flow state data dynamic database module 103.
[0063] The module 101 for constructing a universal local aerodynamic and thermal effect mapping coordinate system and dimensionless local physical quantities is used to construct a dedicated local coordinate system for local areas under different flow states, and adopts a dimensionless method to eliminate the dimensional differences between different data sources. This module enables subsequent deep learning models to identify flow field characteristics at a unified scale and capture nonlinear mapping relationships related to local aerodynamic forces and thermal effects. This module includes two submodules: universal local coordinate system establishment and dimensionless physical quantity. The universal local coordinate system establishment submodule is used to extract local mapping relationships in newly added data, and map local flow fields and aerodynamic and aerodynamic thermal effect data to a universal local coordinate system. The method for establishing the local coordinate system is as follows: Figure 3 This method is based on the assumption that the aerodynamic thermal effect is strongly correlated with the off-wall node, and constructs the following Figure 3The simplified model simplifies the process of aerodynamic thermal effects in the flow field into the effect of the incoming flow on the bottom surface of the microelement, which is defined by the calculation node parameters contained in the microelement. Furthermore, considering the symmetry of the shear stress component within the microelement with respect to the flow parameters, a local coordinate system is established within the microelement through the characteristic velocity V. The characteristic velocity V is generally set as the velocity component v along the shear stress direction of interest at the calculation node near the microelement inlet surface. i The weighted average of:
[0064]
[0065] Among them, n i is the molecular number density or fluid density at the node. Therefore, after the coordinate system and local infinitesimal element are established, the physical quantity dimensionless submodule uses ρ and T near the entrance of the infinitesimal element to dimensionlessly transform the flow parameters:
[0066]
[0067] In the formula, the superscript * represents a dimensionless parameter.
[0068] Deep learning proxy model module 102 uses the data processed by module 101 to train a deep learning model, learning the complex, nonlinear mapping relationships between fluid aerodynamic forces, aerodynamic thermal effects, and the local flow field. The present invention has no specific requirements for the proxy model structure; it only needs to meet the input and output modes, thus providing high flexibility.
[0069] The cross-flow dynamic database module 103 stores the data processed by module 101. Compared to traditional CFD calculation results (often exceeding 1GB for complex working conditions), the data processed by module 101 is much smaller (tens of MB can be used to store thousands of data sets) and is more targeted at gas-solid interactions. When storing data, module 103 uses the local Knudsen number to set a label for the characteristics of the local flow. The local Knudsen number is defined as:
[0070]
[0071] Where U is a physical quantity of the flow field, and the flow field velocity and density are generally used to divide the flow state. Module 103 will detect the flow characteristics (Kn) of the data set used by module 102 before the deep learning model training of module 102, and filter data with similar Kn values from the database to supplement the training data set of module 102, such as Figure 4 The process is shown in Figure 2 to improve the generalization ability of the trained agent model.
[0072] Based on modules 101, 102, and 103, efficient cross-flow aerodynamic thermal effect prediction and database construction can be carried out. The general workflow is as follows:
[0073] In step 201, high-precision CFD simulations and experiments are used to collect aerodynamic and thermal effect data for each region of the flow field. This step ensures that the data possesses sufficient physical realism and fine resolution, providing a solid data foundation for subsequent deep learning model training.
[0074] In step 202, a universal local coordinate system is established for the local regions under different flow regimes. Local dimensionless processing is used to eliminate scale differences between different data sources. This step enables the subsequent deep learning model to identify flow field characteristics at a unified scale and capture the nonlinear mapping relationship between local aerodynamic forces and thermal effects.
[0075] Step 203 calculates the Kn value of local flow data, categorizes and stores the data by flow state, and constructs a cross-flow state data management database. This database not only facilitates subsequent agent model training but also dynamically filters historical data for data similar to the current operating conditions during model operation, thus enhancing data.
[0076] In step 204, based on data preprocessing and local feature extraction, a deep neural network is used to map and model local aerodynamic forces and thermal effects. During the construction of the proxy model, a small amount of data is first analyzed, and then more similar flow pattern data is extracted from the database for data augmentation and refined training to ensure that the model has sufficient predictive capabilities for new flow patterns.
[0077] In step 205, the trained deep learning model is used to predict local aerodynamic and thermal effects in real-time or near-real-time. The prediction results are not only used for engineering design and optimization, but also feed back into CFD calculations through a feedback mechanism, continuously correcting gas-solid interaction parameters during numerical simulations and improving overall simulation accuracy and efficiency.
[0078] Figure 4 The role of module 103 in steps 203 and 204 is further explained, and the workflow is as follows:
[0079] Step 401: Establish a universal local coordinate system, perform local dimensionless processing, and complete the preprocessing of the newly added data.
[0080] Step 402: Use the pre-processed new data to construct a deep learning model training data set.
[0081] Steps 401 to 402 are the usual operations for training a deep learning proxy model. Next, steps 403 to 404 are completed in a dynamic database of cross-flow data to achieve data enhancement.
[0082] Step 403: store the newly added data into the cross-flow state data dynamic database and calculate the local flow field characteristics (local Kn number).
[0083] Step 404: Filter local flow field data outputs with similar flow field characteristics in the database and add them to the deep learning model training data set to achieve data enhancement.
[0084] In step 405, the deep learning agent model is trained using the enhanced training data set to learn the complex nonlinear mapping relationship between the local flow field and the aerodynamic and aerothermal effects.
[0085] Summarize:
[0086] Data-driven deep learning and database construction method: A data-driven deep learning and database construction method is proposed. Through high-precision CFD simulation and experimental data collection, after local coordinate system construction and dimensionless processing, deep learning model training is used to efficiently construct a proxy model of local area cross-flow aerodynamic and thermal effects, realizing the learning of the nonlinear mapping relationship between local aerodynamic and thermal effects.
[0087] Universal local coordinate system and dimensionless processing method: For local areas with different flow states, a universal local coordinate system is constructed, and the coordinate system is established based on the characteristic velocity. The density and temperature near the micro-element inlet are used to non-dimensionalize the flow parameters, eliminating the dimensional differences of multi-source data, unifying the data scale, and improving the model's adaptability to different flow states.
[0088] Construction and management of a dynamic database for cross-flow state data: This system stores processed data, labels local flow features using the local Knudsen number (Kn), and filters similar data based on flow state similarity to supplement the training set, addressing data scarcity issues and improving the efficiency and accuracy of proxy model training.
[0089] Advantages compared with existing technologies:
[0090] Ability to deeply mine local information: By constructing a local coordinate system and making data dimensionless, it can accurately capture local flow field details (such as boundary layer and shock wave area characteristics), improve the accuracy of local aerodynamic and thermal effect predictions, and solve the problem of rough local feature extraction in existing methods.
[0091] Multi-source data fusion capability: The universal local coordinate system and dimensionless processing are generalizable and can seamlessly integrate multi-source data such as CFD simulation and experiments, reduce data noise and inconsistency, and enhance the robustness of the model in cross-flow scenarios (hypersonic reentry, plume contamination, etc.).
[0092] Efficient data utilization: The dynamic database uses the Kn number to filter similar historical data, enabling data enhancement and online model correction. This avoids resource waste in large-scale CFD simulations under new working conditions and improves data utilization efficiency and model generalization capabilities.
[0093] Mutual feedback mechanism between surrogate model and CFD: The prediction results of the surrogate model can feed back into the CFD calculation to correct the gas-solid coupling parameters, forming a "prediction-correction-optimization" closed loop, improving the overall numerical simulation accuracy and efficiency, and realizing the complementary advantages of deep learning and traditional CFD.
[0094] This method constructs an efficient, accurate and adaptable cross-flow aerodynamic thermal effect prediction system, breaking through the bottlenecks of low computational efficiency of traditional CFD and insufficient data utilization of existing deep learning methods. It provides a new solution for engineering design and optimization under complex flow conditions, combining high-precision prediction and real-time response capabilities, significantly reducing computing resource consumption, and promoting the practicality and efficiency of aerodynamic thermal effect prediction technology.
Claims
1. An efficient cross-flow aerodynamic thermal effect prediction and database construction method based on data-driven deep learning, characterized by: The following steps are involved: Through high-precision CFD simulation and experiments, the aerodynamic and thermal effect data of each area in the flow field are collected; For local areas under different flow regimes, a universal local coordinate system is established, and the scale differences between different data sources are eliminated through local dimensionless processing; Calculate the local Knudsen number Kn of local flow field data, classify and store the data according to flow state, and build a dynamic database of cross-flow state data; Based on data preprocessing and local feature extraction, a deep neural network is used to model the local aerodynamic and thermal effect mapping and train a deep learning agent model. The trained deep learning model is used for real-time or quasi-real-time prediction of local aerodynamic and thermal effects. The prediction results are fed back to the CFD calculation through a feedback mechanism to correct the gas-solid interaction parameters.
2. The method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning according to claim 1, characterized in that: A universal local coordinate system is established. Specifically, based on the assumption that the aerodynamic thermal effect is strongly correlated with the off-wall node, a simplified model is constructed to simplify the process of the flow field generating the aerodynamic thermal effect into the effect of the incoming flow on the bottom surface position of the microelement within a microelement, which is defined by the calculation node parameters contained in the microelement; a local coordinate system is established within the microelement through the characteristic velocity V, which is the velocity component v along the shear stress direction of the calculation node near the microelement inlet surface. i The weighted average of is calculated as: Among them, n i is the molecular number density or fluid density at the node.
3. The method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning according to claim 2, characterized in that: The local dimensionless treatment is as follows: the density ρ and temperature T near the microelement inlet are used to dimensionlessly transform the flow parameters. The formula is: ρ * =log 10 ρ / ρ0,p * =log 10 p / p0,T * =log 10 T / T0; where the superscript * represents a dimensionless parameter.
4. The method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning according to claim 1, characterized in that: When storing data in the dynamic database of cross-flow state data, the local Knudsen number Kn is used to set a label for the characteristics of the local flow. The calculation formula of the local Knudsen number Kn is: Where λ is the molecular mean free path, U is the flow field physical quantity, and the flow field velocity and density are used to divide the flow state.
5. The method for predicting and building a database of efficient cross-flow aerodynamic thermal effects based on data-driven deep learning according to claim 1, characterized in that: When training the deep learning agent model, the cross-flow data dynamic database detects the flow characteristics Kn of the data set used, and filters and provides data with similar Kn values from the database to supplement the training data set to achieve data enhancement.
6. The method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning according to claim 5, characterized in that: The specific process of data enhancement is as follows: establish a universal local coordinate system, perform local dimensionless processing, and complete the preprocessing of new data; use the preprocessed new data to construct a deep learning model training data set; store the new data in a dynamic database of cross-flow state data, and calculate the local flow field characteristics Kn; screen the local flow field data output with similar flow field characteristics in the database and add them to the deep learning model training data set; the deep learning agent model uses the enhanced training data set to complete training and learn the complex nonlinear mapping relationship between the local flow field and aerodynamic and aerothermal effects.
7. The method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning according to claim 1, characterized in that: Construct and train a deep learning agent model to learn the complex nonlinear mapping relationship between fluid aerodynamics, aerodynamic thermal effects and local flow fields. There are no special requirements for the agent model structure used, as long as the input and output modes are met.
8. The method for efficiently predicting cross-flow aerodynamic thermal effects and building a database based on data-driven deep learning according to claim 1, characterized in that: Feedback is provided to CFD calculations through a feedback mechanism. Specifically, the prediction results of the deep learning model are used to correct the gas-solid coupling process in the CFD calculation, thereby improving the overall numerical simulation accuracy and efficiency.