Abstract

Weather and climate models rely on parametrisation schemes to represent sub-grid processes that cannot be explicitly resolved on the computational grid. Many such traditional schemes depend on fixed coefficients that are only weakly constrained and tuned offline, often locking in persistent biases and limiting adaptability across regimes, resolutions, and climates.

In this talk, I present a strategy that reframes part of parametrisation design as a sequential control problem by embedding a reinforcement learning (RL) agent within the model, allowing it to observe the evolving state and update selected tunable components online during integration. We evaluate this approach across a hierarchy of idealised environments, from a simple single-parameter bias-correction setting to multi-parameter zonal energy balance models (EBMs), exploring both single-agent and federated multi-agent configurations, the latter mirroring the spatial decomposition used in general circulation models.

Across these settings, we find RL-assisted parameter updates consistently reduce area-weighted RMSE relative to static tuning, with the largest gains emerging in tropical and mid-latitude bands, while federated training accelerates convergence and enables geographically specialised control without sacrificing physically meaningful parameter adjustments. Overall, results from these idealised setups suggest that RL provides a viable pathway toward regime-aware, state-dependent parametrisations and a scalable framework for online learning within numerical weather and climate models.

NOTE – ZOOM ONLINE TALK More information including the Zoom link can be found on the ICCS website