Machine-learning-based Reduced-Order Modelling of a morphing wing in turbulent flow using Proper Orthogonal Decomposition and mesh deformation

• Machine Learning novel methods for turbulent flows predictions. • The application of the Proper Orthogonal Decomposition (POD) based on mesh deformation for morphing wings. • Coupling of the POD based mesh with Long Short-Term Memory (LSTM) models for unsteady flow with wake dynamics. This study introduces a novel approach to predict the turbulent flow dynamics around morphing wings, specifically focusing on the flow around an Airbus A320 airfoil at a Reynolds number of 1 million and an angle of attack (10°). The flow exhibits detachment near 70% of the chord. The proposed method uses an extended Proper Orthogonal Decomposition (POD) mesh-based approach, accounting for real-time wing deformation according to a Travelling Wave (TW) to manipulate flow separation and wake instabilities. The POD is coupled with Machine Learning models, particularly Long Short-Term Memory (LSTM), allowing for extension of the Hi-Fi solution over a longer physical time rapidly. The results are compared to the High-Fidelity approach. The relative error is less than 3% located further downstream in the wake. The study compares several model architectures such as standard and Bidirectional LSTM, ultimately proposing a Hybrid model associating a POD reconstruction to the best LSTM architecture with both models that demonstrates high accuracy in predicting time averaged and unsteady velocity fields based in the wake compared to the other models.