Encoded data assimilation for real-time flow and load estimation in strongly disturbed aerodynamics

Real-time estimation of aerodynamic states is critical for robust control of air vehicles operating in unsteady, disturbed environments. In such regimes, rapidly evolving flow structures are only partially observed through sparse and noisy surface pressure measurements, posing a significant challenge for traditional estimation techniques. This work presents a hybrid data assimilation framework that combines deep learning-based surrogate modeling with low-rank ensemble filtering to reconstruct unsteady flow fields and aerodynamic loads under strong disturbances.

A nonlinear autoencoder is first employed to compress high-dimensional flow states into a compact latent representation that preserves essential flow features. Surrogate models for the latent dynamics and observation operator---the key components of sequential filtering—--are then learned via neural networks and embedded within a low-rank Ensemble Kalman Filter (LR-EnKF). This architecture enables real-time, uncertainty-aware state estimation with significantly reduced computational cost.

The proposed framework is evaluated on gust-encounter simulations over a NACA airfoil across multiple angles of attack. Results show accurate reconstruction of lift and dominant vortical structures, along with credible uncertainty bounds that reflect both sensor noise and the observability of the latent space. Sensor importance analysis highlights the dominant role of leading-edge sensors in state recovery and demonstrates the estimator’s robustness to sensor dropout. The method generalizes well to previously unseen gust conditions, offering a scalable and efficient foundation for data-driven flow estimation and closed-loop control in complex aerodynamic systems.