New physics-informed AI model masters complex aerodynamics at high angles of attack

Loss of control in-flight remains one of the most significant safety challenges in aviation. These critical incidents often occur when an aircraft enters high angles of attack—extreme flight conditions where the flow over wings breaks down, becoming chaotic and highly unpredictable. Designing robust control systems that can maintain stability or automatically recover from these dangerous states requires highly accurate aerodynamic models. However, engineers have long faced a deadlock: traditional physical models are often too simplified to capture the complex dynamics of a stall, while black box artificial intelligence models lack the reliability and interpretability required for safety-critical systems.

A new study published in the Chinese Journal of Aeronautics （https://doi.org/10.1016/j.cja.2025.103943） tries to break this deadlock. Researchers from Northwestern Polytechnical University and Tsinghua University have developed a novel modeling framework called differential equation model long-short term memory neural network (DEM-LSTM). This hybrid approach integrates the theoretical rigor of differential equations with the nonlinear mapping capabilities of LSTM neural network, offering a solution that is both accurate and physically interpretable.

“Accurately predicting unsteady aerodynamic forces is the key to ensuring flight safety and expanding the operational envelope of modern aircraft,” explains Dr. Wenping Song, co-corresponding author of the study. When an aircraft performs sudden maneuvers or encounters severe turbulence, the aerodynamic forces depend not just on its current position, but on its entire motion history—a phenomenon known as unsteady aerodynamics.

Modeling this is notoriously difficult. Traditional methods rely on expert experience—simplified assumptions that often fail to capture complex, multi-axis coupling effects. Conversely, pure data-driven methods act as black boxes. They can mimic training data perfectly but often fail catastrophically when facing new flight conditions, especially when training data is scarce due to the high cost of wind tunnel testing. The industry urgently needs a model that can learn from limited data without violating the laws of physics.

The focus of this research is the DEM-LSTM framework. Instead of letting a neural network predict the aerodynamic forces directly, the researchers used a traditional expert’s experience-informed differential equation model (DEM) as a skeleton. In this framework, the LSTM neural network functions as the muscle. It is tasked with reconstructing the specific a priori terms—the coefficients and functional relationships—within the differential equation.

“By constraining the AI within a differential equation structure, we ensure the model respects the basic mechanisms of flight dynamics,” says Chengpeng Liu, the lead author. “We allow the data to refine the expert's experience-informed terms, rather than replacing the physical framework entirely.”

The team validated the method using two rigorous test cases: single-degree-of-freedom pitch oscillations of a missile and two-degree-of-freedom coupled motions of a fighter. The results showed that the DEM-LSTM framework achieved significantly higher accuracy and generalization capabilities than differential equation model, LSTM and multi-fidelity neural network model.

While the current results are promising, the transition from lab theory to engineering application requires further refinement. The researchers will continue to study the noise suppression method of the model as well as the extrapolation boundaries. The final goal of this research is to provide the aerospace industry with a standard, high-fidelity modeling tool. By solving the nonlinear unsteady aerodynamic modeling problem, this technology paves the way for the design of next-generation aircraft with superior maneuverability, robustness, and safety.

Original Source

Chengpeng LIU, Wenping SONG, Chenzhou XU, et al. Physics-informed data-driven nonlinear unsteady aerodynamic modelling at high angles of attack[J]. Chinese Journal of Aeronautics, 2026, 39(7): 103943. https://doi.org/10.1016/j.cja.2025.103943

About Chinese Journal of Aeronautics 

Chinese Journal of Aeronautics (CJA) is an open access, peer-reviewed international journal covering all aspects of aerospace engineering, monthly published by Elsevier. The Journal reports the scientific and technological achievements and frontiers in aeronautic engineering and astronautic engineering, in both theory and practice. CJA is indexed in SCI (IF = 6.9, Q1), EI, IAA, AJ, CSA, Scopus.