Inductive Biases for Robust Climate Emulation Across Forecast Timescales

Machine-learning emulators have demonstrated remarkable skill for weather prediction and short-range forecasting, yet their behaviour, as forecasts extend toward seasonal and longer timescales, remains less well explored and understood. Approaching these horizons, forecast skill is shaped less by short-range error growth, while variations in background states or system parameters increasingly influence the evolving dynamics. Understanding if and how different neural architectures perform with such changes is therefore central to assessing their suitability for emulation beyond medium range weather prediction, where robustness plays an increasingly important role. In this work, we investigate how inductive biases encoded in deep-learning architectures influence their ability to represent and evolve dynamics as forecasts move into windows nearing and sometimes beyond their training data.

We use the idealised climate model ISCA as a controlled testbed, enabling systematic variation of planetary parameters and initial conditions while retaining a fixed underlying set of governing equations. Emulators are trained on ensembles of trajectories sampled from a restricted parameter range and evaluated under progressively more challenging ID/OOD settings. This framework allows us to disentangle errors arising from finite-horizon forecasting from those associated with longer-timescale dynamical shifts, providing insight into which architectural biases promote stability, physical consistency, and robustness as machine-learning models are pushed from shorter term prediction toward longer time scale emulations.