📝 SummarXiv

Abstract

Accurately reconstructing and forecasting high-resolution (HR) states from computationally cheap low-resolution (LR) observations is central to estimation-and-control of spatio-temporal PDE systems. We develop a unified superresolution pipeline based on the reduced-order autodifferentiable Ensemble Kalman filter (ROAD-EnKF). The method learns a low-dimensional latent dynamics model and a nonlinear decoder from latent variables to HR fields; the learned pair is embedded in an EnKF, enabling simultaneous state estimation and control-oriented forecasting with quantified uncertainty. We evaluate on three benchmarks: 1-D viscous Burgers equation (shock formation), Kuramoto-Sivashinsky (KS) equation (chaotic dynamics), and 2-D Navier-Stokes-Kraichnan turbulence (NSKT) (vortex decaying dynamics at Re 16,000). LR data are obtained by factors of 4-8 downsampling per spatial dimension and are corrupted with noise. On Burgers and KS, the latent models remain stable far beyond the observation window, accurately predicting shock propagation and chaotic attractor statistics up to 150 steps. On 2-D NSKT, the approach preserves the kinetic-energy spectrum and enstrophy budget of the HR data, indicating suitability for control scenarios that depend on fine-scale flow features. These results position ROAD-EnKF as a principled and efficient framework for physics-constrained superresolution, bridging LR sensing and HR actuation across diverse PDE regimes.