Demystifying Data-Driven Probabilistic Medium-Range Weather Forecasting

TL;DR

The paper addresses fragmentation in data-driven probabilistic medium-range weather forecasting by introducing ATLAS, a simple yet scalable latent-space transformer framework that decouples global atmospheric evolution from high-resolution physics. ATLAS operates in a directly downsampled latent space and uses a history-conditioned local projector, and is compatible with stochastic interpolants, diffusion models, and CRPS-based ensemble training, demonstrating state-of-the-art probabilistic skill against IFS and GenCast. Key contributions include a unified latent modeling approach, method-agnostic backbone applicability, and robust performance across a range of probabilistic estimators, along with extensive validation on ERA5 data and tropical cyclone tracking. The work suggests that scaling a general-purpose foundation-like model, rather than engineering bespoke architectures or recipes, yields substantial forecast improvements with practical benefits for fast, ensemble-based predictions.

Abstract

The recent revolution in data-driven methods for weather forecasting has lead to a fragmented landscape of complex, bespoke architectures and training strategies, obscuring the fundamental drivers of forecast accuracy. Here, we demonstrate that state-of-the-art probabilistic skill requires neither intricate architectural constraints nor specialized training heuristics. We introduce a scalable framework for learning multi-scale atmospheric dynamics by combining a directly downsampled latent space with a history-conditioned local projector that resolves high-resolution physics. We find that our framework design is robust to the choice of probabilistic estimator, seamlessly supporting stochastic interpolants, diffusion models, and CRPS-based ensemble training. Validated against the Integrated Forecasting System and the deep learning probabilistic model GenCast, our framework achieves statistically significant improvements on most of the variables. These results suggest scaling a general-purpose model is sufficient for state-of-the-art medium-range prediction, eliminating the need for tailored training recipes and proving effective across the full spectrum of probabilistic frameworks.

Figures (12)

• Figure 1: Scorecard comparison of ATLAS-SI vs. IFS-ENS (a) and GenCast (b) for a fifteen day forecast with 56 ensemble members averaged over ERA5 initial conditions in 2020. Left shows percent improvement in the RMSE of the ensemble mean and right in the ensemble CRPS.
• Figure 2: Overview of our approach: Atmospheric states $x_0$ and $x_{-1}$ are first encoded into a latent representations $z_0$ and $z_{-1}$. These are mapped through a probabilistic model with a transformer backbone into a latent representation of residuals $r_1$ of the next time step. These residuals are deterministically decoded back into original space with a transformer-based decoder which is also conditioned on the high resolution initial state. Note that the probabilistic model uses a latent DiT with global attention, while the decoder relies on a local attention DiT.
• Figure 3: Heatmaps comparing Atlas and IFS across RMSE and CRPS, for all three variants, SI, EDM and CRPS.
• Figure 4: Comparing Atlas and GenCast across RMSE and CRPS with statistical significance (p<0.05), for all three variants, SI, EDM and CRPS. Green means Atlas is better, red means GenCast is better, white means the difference is not statistically significant.
• Figure 5: 15-day rollout comparison between ATLAS, IFS, and GenCast.