Core Concepts
Discovering interpretable, closed-form equations for subgrid-scale closures is crucial for Earth system modeling.
Abstract
The manuscript discusses the challenges and promises of learning closed-form equations for subgrid-scale closures using high-fidelity data. It highlights the importance of physics-informed loss functions, libraries, metrics, and sparsity selections to achieve accurate and stable closures. The study focuses on 2D turbulence/convection simulations to robustly discover closures for momentum and heat fluxes. Physics-based parameterizations in climate models are critiqued for their shortcomings, leading to biases and uncertainties. The paper explores equation-discovery techniques like relevance vector machine (RVM) to learn interpretable closures directly from simulations. It emphasizes the need for interpretability, generalizability, computational cost considerations in closure modeling efforts.
Stats
In 2D-FHIT cases K1-K3, the average correlation coefficient (CC) values of the discovered closure are around 0.99.
For RBC cases R1-R3, the CC values of the discovered closure also demonstrate high accuracy with values above 0.95.