The high maneuverability of modern fighters in close air combat imposes significant cognitive demands on pilots, making rapid, accurate decision-making challenging. While reinforcement learning (RL) has shown promise in this domain, the existing methods often lack strategic depth and generalization in complex, high-dimensional environments. To address these limitations, this paper proposes an optimized self-play method enhanced by advancements in fighter modeling, neural network design, and algorithmic frameworks. This study employs a six-degree-of-freedom (6-DOF) F-16 fighter model based on open-source aerodynamic data, featuring airborne equipment and a realistic visual simulation platform, unlike traditional 3-DOF models. To capture temporal dynamics, Long Short-Term Memory (LSTM) layers are integrated into the neural network, complemented by delayed input stacking. The RL environment incorporates expert strategies, curiosity-driven rewards, and curriculum learning to improve adaptability and strategic decision-making. Experimental results demonstrate that the proposed approach achieves a winning rate exceeding 90% against classical single-agent methods. Additionally, through enhanced 3D visual platforms, we conducted human-agent confrontation experiments, where the agent attained an average winning rate of over 75%. The agent’s maneuver trajectories closely align with human pilot strategies, showcasing its potential in decision-making and pilot training applications. This study highlights the effectiveness of integrating advanced modeling and self-play techniques in developing robust air combat decision-making systems.

Aerodynamic noise originates from pressure fluctuations during gas flow, which can lead to acoustic fatigue and acoustic-structural coupling, and is a significant factor affecting the safety and comfort of aircraft. Research methods for aerodynamic noise primarily include theoretical approaches, wind tunnel testing, and numerical simulation. However, these methods suffer from limitations such as singular measurement results, difficulty in establishing effective correlations with flow structures, and challenges in obtaining high-precision noise data. Machine learning methods, characterized by their efficiency, speed, and low cost, have shown great potential in the field of aeronautical aerodynamic noise. This paper provides an overview of the latest research progress in machine learning applied to aeronautical aerodynamic noise, with a focus on the reconstruction of sound fields under sparse measurement points and the prediction of aerodynamic noise. Finally, the paper analyzes common issues in machine learning methods for aerodynamic noise research, such as weak generalizability, insufficient prediction accuracy, and lack of physical interpretability, and looks forward to future development trends, offering a reference for aerodynamic noise research based on machine learning methods.

To enhance the independent development capability and efficiency of new strategic aircraft, it is crucial to accurately acquire aerodynamic data over a wide range of high Reynolds numbers. There is an urgent need to improve the accuracy of aerodynamic simulations for complex flow conditions at flight Reynolds numbers, and to address the high cost associated with ground testing in cryogenic wind tunnels. The experimental data at high Reynolds numbers are few and sparse, and there are serious data imbalance and few-shot problems in the accurate aerodynamic modeling under wide-range Reynolds numbers. To solve the contradiction between the cost and accuracy of the aerodynamic model under variable Reynolds number, and focus on the cost reduction and efficiency increase of subsequent aircraft design, this study takes the CHN-T1 transport aircraft as the research objective. Based on data fusion and information transfer techniques, we achieve the rapid prediction of variable Reynolds number aerodynamics through few-shot learning. This approach reduces the reliance on high Reynolds number samples in the modeling process. In the present study, a wide-ranging variable Reynolds number aerodynamic dataset is constructed using a combination of 18 aerodynamic curves. These curves encompass sub-transonic speeds and variable Reynolds numbers ranging from millions to tens of millions. Additionally, various complexity cases are designed for demonstration. As a verification, a benchmark is established using high-fidelity experimental data through a single source method. The characteristics of different methods are then compared. The results demonstrate that when approximately 10 high-fidelity aerodynamic curves are utilized as modeling data, the data fusion neural network reduces the root mean square error of aerodynamic modeling by over 50%. Additionally, the information transfer method reduces the error by at least 40%.

Aerodynamic surrogate modeling mostly relies only on integrated loads data obtained from simulation or experiment, while neglecting and wasting the valuable distributed physical information on the surface. To make full use of both integrated and distributed loads, a modeling paradigm, called the heterogeneous data-driven aerodynamic modeling, is presented. The essential concept is to incorporate the physical information of distributed loads as additional constraints within the end-to-end aerodynamic modeling. Towards heterogenous data, a novel and easily applicable physical feature embedding modeling framework is designed. This framework extracts low-dimensional physical features from pressure distribution and then effectively enhances the modeling of the integrated loads via feature embedding. The proposed framework can be coupled with multiple feature extraction methods, and the well-performed generalization capabilities over different airfoils are verified through a transonic case. Compared with traditional direct modeling, the proposed framework can reduce testing errors by almost 50%. Given the same prediction accuracy, it can save more than half of the training samples. Furthermore, the visualization analysis has revealed a significant correlation between the discovered low-dimensional physical features and the heterogeneous aerodynamic loads, which shows the interpretability and credibility of the superior performance offered by the proposed deep learning framework.

Air combat is an important element in the three-dimensional nature of war, and intelligent air combat has become a hotspot and focus of research in the military field both domestically and internationally. Deep reinforcement learning is an important technological approach to achieving air combat intelligence. To address the challenge of constructing high-level opponents in single agent training method, a self-play based air combat agent training method is proposed, and a visualization research platform is built to develop a decision-making agent for close-range air combat. The field knowledge of pilots is embedded in the design process of the agent’s observation, action, and reward, training the agent to convergence. Simulation experiments show that the air combat tactics of agent gradually improves by self-play training, achieving a win rate of over 70% against the decision making by single agent training and the emerging of the strategies similar to human “single/double loop” tactics.

Aeroheating prediction with high efficiency and high accuracy is crucial for the design of hypersonic vehicles. However, the increasing shape complexity and tight design period of hypersonic vehicles make it difficult for existing methods to meet the requirements of efficient and accurate aeroheating prediction. In this study, a localized data-driven modeling method for rapid aeroheating prediction is developed based on the boundary layer theory and the support vector machine. Firstly, the outer edge boundary layer information is obtained by solving the Euler equations, and the RANS method is used to generate samples of heat flux distributions. Then, a feature selection approach is developed to acquire the outer edge boundary layer features. Finally, the support vector machine is used to construct the aeroheating prediction model to achieve the mapping between the outer edge boundary layer features and the heat flux on the wall. Results of the aeroheating prediction for a double ellipsoid and a two-stage compression surface show that the model considers local boundary conditions such as the non-uniform wall temperature, and has high accuracy as well as good extrapolation and generalization capability. The relative errors of heat flux between the model prediction and the RANS calculation are less than 5%. Moreover, for the aeroheating flux prediction along the center line on the upper surface of the double ellipsoid, the prediction ability of the present model is better than the traditional proper orthogonal decomposition (POD) reduction method, especially the prediction accuracy of the present model in the extrapolation regime is more than four times higher than that of the POD reduction model.

Aerodynamic analysis of aircraft design often requires a large amount of high-fidelity (HF) aerodynamic data to improve the performance of aircraft design. However, the acquisition cost is very high. In order to alleviate the contradiction between modeling cost and accuracy, this paper constructs a multi-fidelity aerodynamic data fusion model by associating data with different fidelity. Furthermore, an optimal correlation point selection method and a uniformly enhanced sequential sampling method are proposed to achieve the efficient initialization and fastest convergence of variable-fidelity models based on co-Kriging. As a validation, standard numerical examples are selected to carry out modeling study, and the accuracy of the method is checked by comparing the statistical variables. Finally, the framework is successfully applied in the transonic aerodynamic engineering case of the NACA0012 airfoil. The results show that compared with the traditional model, the proposed method can greatly improve the convergence accuracy and modeling efficiency of the variable-fidelity model with only a small number of high-fidelity samples, which effectively reduces the sampling cost. Compared to the high-fidelity single precision sequence modeling, the error can be reduced by more than a half.

Numerical simulations in recent years have shown that vortex-induced vibrations (VIV) of cylinders can occur at subcritical Reynolds numbers as low as 20, but relevant experimental evidence for the existence of such phenomenon has not been observed. In this paper, we first built a rotating channel for low-Reynolds-number experiments. Then, VIV at subcritical Reynolds numbers were studied in this rotating channel. The effects of support stiffness and Reynolds number on the VIV were investigated. The lowest Reynolds number for the occurence of VIV of a cylinder is 23, which is close to the numerical simulation results. It confirms that the subcritical VIV indeed exists. In addition, von Kármán vortex shedding was found during VIV and the vortex shedding frequency is the same as that of the cylinder. It indicates that the elastic support makes the flow less stable, which also agrees well with previous numerical results.

Wind tunnel and flight tests are two of the most important methods for aerodynamic analyses and optimization design during the development of aircraft. However, under hypersonic flight conditions, the real gas effects, viscous interference effects, and multi-scale flow fields pose huge challenges to aerodynamic prediction. To improve the consistency of aerodynamic data between wind-tunnel and flight tests, an aerodynamic data fusion framework based on the random forest method for data mining is proposed and applied in the aerodynamic data fusion of a hypersonic aircraft. Feature analyses and ranking of aerodynamic data obtained by ground wind tunnel tests are conducted first. Then aerodynamic data in a flight envelop are cross-validated. Results show that the machine learning framework based on the random forest has good prediction and extrapolation capabilities for the correlation of aerodynamic data obtained by wind-tunnel and flight tests, and can effectively improve the prediction accuracy of aerodynamic data. The method provides a promising solution to the multi-source fusion of aerodynamic data in complex environments.