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

TL;DR

A multi-variable Vision Transformer architecture with a shared encoder and variable-specific decoders (1EMD) that consistently outperforms alternative multi-variable baselines and substantially reduces computational cost.

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) address this limitation through dynamical downscaling, albeit at considerable computational cost and with limited flexibility. Deep learning has emerged as an efficient data-driven alternative; however, most existing approaches focus on single-variable models that downscale one variable at a time. This paradigm can lead to redundant computation, limited contextual awareness, and weak cross-variable interactions.To address these limitations, we propose a multi-variable Vision Transformer (ViT) architecture with a shared encoder and variable-specific decoders (1EMD). The proposed model jointly predicts six key climate variables: surface temperature, wind speed, 500 hPa geopotential height, total precipitation, surface downwelling shortwave radiation, and surface downwelling longwave radiation, directly from GCM-resolution inputs, emulating RCM-scale downscaling over Europe. Compared to single-variable ViT models, the 1EMD architecture improves performance across all six variables, achieving an average MSE reduction of approximately 5.5% under a fair and controlled comparison. It also consistently outperforms alternative multi-variable baselines, including a single-decoder ViT and a multi-variable U-Net. Moreover, multi-variable models substantially reduce computational cost, yielding a 29-32% lower inference time per variable compared to single-variable approaches. Overall, our results demonstrate that multi-variable modeling provides systematic advantages for high-resolution climate downscaling in terms of both accuracy and efficiency. Among the evaluated architectures, the proposed 1EMD ViT achieves the most favorable trade-off between predictive performance and computational cost.

Figures (8)

• Figure 1: Super-resolution problem setup. A low-resolution field (LR, left) is fed into a neural model (center), which outputs a super-resolved field (HR, right).
• Figure 2: Visualization of a single daily sample for the six studied variables: Surface Temperature, Wind Speed, Geopotential Height, Total Precipitation, Surface Downwelling Shortwave Radiation and Surface Downwelling Longwave Radiation. Each variable is shown for both the high-resolution Regional Climate Model ICTP-RegCM4-6 (top row) and the low-resolution Global Climate Model MPI-ESM-LR (bottom row) re-gridded onto the same European domain.
• Figure 3: Comparison between a single-variable encoder-decoder model and multi-variable models using either a single shared decoder (1E1D) or multiple decoders branching from a shared encoder (1EMD).
• Figure 4: Architecture of the multi-variable single-encoder multi-decoder model (1EMD). Concatenated input variables are divided into patches, embedded with positional encodings, and processed as a sequence by six transformer blocks, each consisting of multi-head self-attention (6 heads) followed by an MLP. The output sequence is then reshaped into a spatial grid and passed to separate CNN-based decoders, each composed of three upsampling stages interleaved with residual blocks.
• Figure 5: Comparison of the examined models (Single var ViT, multi var ViT 1E1D, Multi var ViT 1EMD), target RCM, and bilinearly interpolated GCM for a sample of the test set. For each sample, each row contains Geopotential, Wind speed, Temperature, Total Precipitation, Surface Downwelling Shortwave Radiation and Surface Downwelling Longwave Radiation for the same timestamp.