In this work , led by Andrew Ross, with Ziwei Li, Pavel Perezhogin, Carlos Fernandez-Granda and Laure Zanna, we provide a framework for systematically benchmarking the offline and online performance of physical and ML-based subgrid parameterizations. We find that the choice of filtering operator is critical for performance. To help with interpretability, we also propose a novel equation-discovery approach combining linear regression and genetic programming which generalizes better than physical and neural network parameterizations.

See code and notebooks here: https://github.com/m2lines/pyqg_parameterization_benchmarks