Feature-preserving Latent-EnKF for Data Assimilation of Flows with Shocks
Standard Ensemble Kalman Filters fail for flows with shocks because they produce spurious oscillations when uncertainty is in shock location. This method introduces a feature-preserving latent-EnKF that performs data assimilation in a learned low-dimensional latent space. It uses an auto-decoder to encode sharp flow features as smooth variations, allowing the filter to translate shocks coherently instead of superposing them.