ML Microphysics Scheme Achieves Stable Decade-Scale ICON Runs Without Two Tuning Parameters

The team trained a classifier to detect active microphysical grid cells and a regressor to predict tendencies on global convection-permitting output. Physical constraints enforcing mass positivity and preventing overshoots proved decisive for stable online coupling; without them the hybrid model crashed within days despite strong offline metrics. Dataset curation that exposed the network to a wider range of coarse-scale states further reduced drift.

In decade-long AMIP-style runs the ML scheme reproduced global cloud and precipitation climatologies at levels statistically indistinguishable from the default graupel parameterization. It eliminated two microphysics-specific tuning parameters yet did not reduce long-standing mean-state biases in liquid water path or shortwave cloud forcing. The result demonstrates that offline skill alone is insufficient; stability requires explicit physical guardrails.

This advance directly targets the cloud-related uncertainty that limits sub-seasonal to decadal forecast skill. Because ICON is used operationally for both numerical weather prediction and climate projection, the hybrid microphysics opens a path to reduced ensemble spread in precipitation and radiation forecasts within the next annual cycle.

Next steps include embedding the scheme in coupled ocean-atmosphere configurations and testing whether targeted fine-tuning on bias-sensitive regimes can finally shrink climatological errors that the current implementation leaves untouched.