Graph neural networks based climate emulator for kilometer scale hourly precipitation : a novel hybrid imperfect approach

Regional climate emulators provide computationally efficient tools for generating high-resolution climate projections, bridging the gap between coarse-scale models and the detailed resolution required for local-scale hazard assessments. Climate hazards from extreme precipitation events are projected to increase in frequency and intensity under global warming, emphasizing the need for accurate modeling of convective processes. However, traditional numerical methods are constrained by low resolution or the high computational costs of kilometer-scale simulations.

To overcome these limitations, we introduce GNN4CD, a novel deep learning emulator that estimates kilometer-scale (3 km) hourly precipitation from coarse atmospheric data (~25 km). The model leverages graph neural networks and a hybrid imperfect approach (HIA) for downscaling, initially trained on ERA5 reanalysis and observational data, and applied using regional climate model (RegCM) data for present-day and future projections.

GNN4CD demonstrates exceptional performance in reproducing precipitation distributions, seasonal diurnal cycles, and extreme percentiles across Italy, even when trained on northern Italy alone. The model captures shifts in precipitation distributions, particularly for extremes, across historical, mid-century, and end-of-century scenarios. Additionally, evaluations using an ensemble of convection-permitting regional models confirm GNN4CD's ability to replicate ensemble spreads for both present-day and future projections essential for estimating the uncertainty in the future climate change signal..