Vision Transformers for Multi-Variable Climate Downscaling: Emulating Regional Climate Models with a Shared Encoder and Multi-Decoder Architecture

Abstract: Global Climate Models (GCMs) are critical for simulating large-scale climate dynamics, but their coarse spatial resolution limits their applicability in regional studies. Regional Climate Models (RCMs) refine this through dynamic downscaling, albeit at considerable computational cost and with limited flexibility. While deep learning has emerged as an efficient data-driven alternative, most existing studies have focused on single-variable models that downscale one variable at a time. This approach can lead to limited contextual awareness, redundant computation, and lack of cross-variable interaction. Our study addresses these limitations by proposing a multi-task, multi-variable Vision Transformer (ViT) architecture with a shared encoder and variable-specific decoders (1EMD). The proposed architecture jointly predicts three key climate variables: surface temperature (tas), wind speed (sfcWind), and 500 hPa geopotential height (zg500), directly from GCM-resolution inputs, emulating RCM-scale downscaling over Europe. We show that our multi-variable approach achieves positive cross-variable knowledge transfer and consistently outperforms single-variable baselines trained under identical conditions, while also improving computational efficiency. These results demonstrate the effectiveness of multi-variable modeling for high-resolution climate downscaling.