Hybrid Physics-ML Framework Reduces Compute for Quasi-Geostrophic Parameter Tuning by 60 Percent

The study applies a two-stage pipeline to subgrid parameterization in a quasi-geostrophic turbulence model. Green's Function Optimization first ranks normalized parameter sensitivities and prunes the active space; nonlinear surrogates then screen candidates with short 30-day runs before expensive 180-day verification. Across seven strategies the hybrid pipelines reached practical saturation faster than standalone GP or NN baselines that required 7740 and 6660 simulation-days. Because screening, dimensionality reduction, and fidelity scheduling vary together, the design cannot isolate which step drove the largest efficiency gain.

Earth system models face the same expensive nonlinear interactions when tuning convection and cloud schemes that dominate uncertainty in CMIP projections. The reported sample-efficiency advantage aligns with earlier multi-fidelity work on Gaussian processes for aerodynamic design, suggesting the approach could transfer if the initial sensitivity ranking remains stable under realistic forcing. Full-scale testing on GCMs with interactive chemistry and ocean coupling will be required before policy-relevant parameter sets are updated.

The main limitation is the proof-of-concept scope: results rest on a single idealized model and a modest ensemble size. Larger ensembles and out-of-sample verification on held-out initial conditions would strengthen claims that the observed 64-65 percent gains generalize beyond the QG setting.