PhD Thesis Presentation - Deep Learning-Based Subgrid Parameterizations for Unresolved Physics

Deep learning (DL)-based parameterization schemes have shown great potential for weather and climate forecasts in recent years. However, their high computational cost remains a major obstacle to operational implementation. In parallel, as numerical weather prediction models move toward kilometer-scale resolutions, the accurate representation of subgrid-scale (SGS) turbulence in the so-called gray zone has emerged as another critical challenge. To address these two issues, this thesis conducts two complementary studies.

The first work focuses on improving the efficiency of DL-based parameterizations while minimizing any negative impact on their efficacy. We propose three methods, reducing the amount of model parameters, focusing on critical sub-domain information only as the input, or combining both. In addition, the loss function is redesigned and the baseline model's forecast lead time is extended. Through these methods and using the solver-in-the-loop training approach, we reduced the model’s computing time while preserving its prediction skills.

The second study addresses the gray-zone turbulence problem by developing a novel large-eddy simulation (LES) framework using the Python library JAX. The new LES code allows GPU acceleration and automatic differentiation. A simple autoencoder (AE) model is trained as the DL-based SGS model. Applied to a classic warm bubble case, the physics-deep learning hybrid model can accurately simulate the expansion of the thermal bubble and the development of rotors surrounding the center of the bubble at a gray-zone resolution. The gray-zone simulation results are comparable to those of the benchmark LES resolution.

Overall, this thesis presents technical solutions for model efficiency improvement and a new LES framework, addressing two key challenges in data-driven parameterization models.