Title: U-Cast: An Efficient and Surprisingly Straightforward Frontier in Probabilistic AI Weather Forecasting

Abstract:

While artificial intelligence-driven weather prediction has achieved parity with traditional physics-based ensemble methods, the current state-of-the-art (SOTA) approaches typically demand specialized architectures and substantial computational resources, thereby establishing a significant barrier to entry. This study challenges the notion that such complexity is required for top-tier performance. We present U-Cast, a probabilistic forecasting model that leverages a conventional U-Net backbone and a streamlined training protocol. This protocol consists of deterministic pre-training using Mean Absolute Error (MAE), followed by brief probabilistic fine-tuning via the Continuous Ranked Probability Score (CRPS), incorporating Monte Carlo Dropout to introduce stochasticity.

Despite its simplicity, U-Cast achieves probabilistic skill levels that are comparable to or superior than those of GenCast and the Integrated Forecasting System (IFS) ENS at a $1.5^\circ$ resolution. Furthermore, it significantly lowers computational demands: training compute is reduced by more than $10\times$ relative to leading CRPS-based models, and inference latency is cut by over $10\times$ when compared to diffusion-based systems. The entire training process requires fewer than 12 H200 GPU-days, and the model can produce a 15-day ensemble forecast in just 3 seconds. These findings indicate that scalable, general-purpose architectures, combined with efficient training strategies, can rival complex, domain-specific designs at a fraction of the cost, thereby democratizing access to frontier probabilistic weather modeling for the wider research community.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC