Stable Machine Learning based Radiation Emulation for ICON

Katharina Hafner, Fernando Iglesias‐Suarez, Sara Shamekh, Pierre Gentine, M. A. Giorgetta, Robert L. Pincus, Veronika Eyring

Preprint · 12 May 2025 · Columbia University (LEAP, Lamont-Doherty), Max Planck Institute for Meteorology

Paper PDF

In one sentence. An ML radiation emulator in ICON runs four times faster than RRTMGP and stays stable over 10-year AMIP-like runs and +4 K SST warming.

Abstract

To enhance both accuracy and computational efficiency in Earth system models, machine learning (ML)-based parameterizations are increasingly being developed as alternatives to traditional physics-based schemes, particularly for subgrid-scale processes. In the context of radiation parameterization, where computational demands are high, ML-based emulators offer a significant speed advantage, which is crucial for reducing uncertainties associated with infrequent radiation calls. Despite promising offline performance, these ML-based radiation emulators have been minimally tested in online settings. In this study, we implement an ML-based radiation emulator into the ICOsahedral Non-hydrostatic (ICON) model and assess its online performance in an AMIP-like simulation with prescribed sea surface temperatures, sea ice concentrations, and constant greenhouse gas, excluding aerosol forcings. Our results show that the ML-based emulator runs four times faster on CPU compared to the traditional RTE+RRTMGP parameterization using the same radiation time step. The coupled simulation reaches a new equilibrium with a slightly altered cloud distribution, but remains stable for a 10-year period tested here with no increasing divergence from the reference simulations over time. The ML-based radiation scheme even runs stable in an experiment in which the sea surface temperature is increased by $4~K$.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.