Data assimilation for energy-aware hybrid models

Abstract: This work integrates ensemble-based data assimilation (DA) with the energy-aware hybrid modeling approach, applied to a three-layer quasi-geostrophic (QG) model of the Gulf Stream flow. Building on prior DA success in the QG channel regime, where stochastic corrections based on EOFs were effective, we show that this method fails to address persistent errors in the more complex, dynamically richer Gulf Stream setting.To overcome this, we employ a hybrid model that controls energy at selected scales, maintaining dynamic consistency and physical realism. We evaluate the combined effect of hybrid modeling and DA, using a particle filter which combines model reduction, tempering, jittering, and nudging. Numerical experiments show that the hybrid model reproduces both the large-scale jet and small-scale vortices seen in high-resolution reference simulations, but missing in the standard (non-hybrid) QG model. When DA is incorporated, the hybrid model further reduces tracking error and ensemble divergence. Moreover, targeted assimilation from the most energetic region matches tracking error and uncertainty reduction of full-domain networks, highlighting the critical importance of observation network design. These findings demonstrate that combining energy-aware hybrid modeling with ensemble-based DA enables high-fidelity, computationally efficient tracking of the reference solution even under sparse, noisy, localized observations.