Dynamic subgrid-scale (SGS) turbulence parameterizations in Large Eddy Simulation (LES) achieve superior physical fidelity but impose 2–4× computational overhead compared to static schemes, creating a critical bottleneck for high-resolution atmospheric modeling on HPC systems. Neural network based emulation offers a pathway to comparable accuracy at reduced computational cost, but realizing this potential requires architectures that generalize reliably across diverse atmospheric conditions and variable grid configurations. We systematically compare two physics-aware multi-task learning strategies for emulating Smagorinsky-based SGS closure in the UK Met Office NERC Cloud Model (MONC): a baseline approach using Richardson number prediction as auxiliary gradient regularization, and an Ri-conditioned approach that explicitly feeds predicted stability into coefficient (viscosity and diffusion) prediction heads. Evaluating 54 model configurations across three neural architectures (multi-layer perceptron (MLP), MLP with residual blocks (ResMLP) and Tabular Transformer (TabTransformer)) trained on mixed-resolution, multi-regime atmospheric data (66% coarse tropical convection, 34% fine shallow cumulus), we find that uncertainty-based task weighting consistently outperforms manual tuning and dynamic weighting alternatives. The simple MLPs with Richardson conditioning provide the best robustness-accuracy trade-off under distribution shift during inference, and the architectural complexity amplifies cross-regime failures despite improving in-distribution metrics. Notably, models maintain physical constraint compliance even when predictive accuracy degrades substantially, suggesting that the data coverage limitations, rather than any fundamental physics incompatibility, drive the cross-regime transfer failures. All results represent offline validation on static simulation data. Ongoing work focuses on online MONC integration to assess numerical stability, energy conservation, and computational performance under coupled feedback dynamics.
Panda et al. (Mon,) studied this question.