Cascaded score-based emulation of Earth system models for impact evaluation with SCALES-MESH

Today's climate adaptation and mitigation planning tasks require rapid access to large ensembles of climate projections for a wide range of emissions scenarios, including overshoot scenarios. While Earth system models (ESMs) provide physically consistent projections, their high computational cost limits scenario exploration. Climate emulators -  statistical or machine-learning-based models trained on ESM data to generate data replicating the ESMs behaviour for a multitude of emissions scenarios - are therefore proposed to deliver these projections efficiently. Here we present the novel modular SCALES–MESH emulator framework, combining physics-based regional projections with AI downscaling capabilities. The SCALES module translates projections of global mean surface air temperature into regional surface air temperature projections aggregated over the AR6-IPCC regions, while the MESH module performs spatio-temporal downscaling to gridded fields using a conditional score-based generative model. MESH is trained on multiple datasets and evaluated against parent ESMs using spatial, temporal, and distributional diagnostics. Results show that the emulator captures regional patterns, temporal variability, and probability distributions of emulated climate variables, including during warming and cooling phases of overshoot scenarios. We further demonstrate the potential for transfer learning across ESMs, pointing toward scalable multi-model and resolution-agnostic emulation. Together, SCALES–MESH enables rapid, flexible, and physically grounded exploration of climate futures, supporting decision-relevant climate risk assessment at unprecedented scope.