U-Cast: A Surprisingly Simple and Efficient Frontier Probabilistic AI Weather Forecaster

AI-based weather forecasting now rivals traditional physics-based ensembles, but state-of-the-art (SOTA) models rely on specialized architectures and massive computational budgets, creating a high barrier to entry. We demonstrate that such complexity is unnecessary for frontier performance. We introduce U-Cast, a probabilistic forecaster built on a standard U-Net backbone trained with a simple recipe: deterministic pre-training on Mean Absolute Error followed by short probabilistic fine-tuning on the Continuous Ranked Probability Score (CRPS) using Monte Carlo Dropout for stochasticity. As a result, our model matches or exceeds the probabilistic skill of GenCast and IFS ENS at $1.5^\circ$ resolution while reducing training compute by over $10\times$ compared to leading CRPS-based models and inference latency by over $10\times$ compared to diffusion-based models. U-Cast trains in under 12 H200 GPU-days and generates a 15-day ensemble forecast in 3 seconds. These results suggest that scalable, general-purpose architectures paired with efficient training curricula can match complex domain-specific designs at a fraction of the cost, opening the training of frontier probabilistic weather models to the broader community.

Modern AI weather models can now compete with the large physics-based systems used by meteorological centers, but many of the strongest models are technically complex and expensive to train. This makes it hard for smaller research groups to reproduce, improve, or even test new ideas in frontier weather forecasting. We asked: Is all this complexity really necessary? We built U-Cast, a weather forecaster using a standard, off-the-shelf neural network design, called a U-Net. The key insight is a two-stage training recipe: first, teach the network to predict the single most likely future weather, then briefly fine-tune it to generate a range of plausible scenarios that capture forecast uncertainty. To produce these alternative futures, U-Cast uses a simple randomness technique instead of a costly, specialized forecasting system. On standard global weather benchmarks, U-Cast matches leading AI and physics-based forecasting systems, while using much less computing power and producing forecasts in seconds. This suggests that strong probabilistic weather forecasting can be made more accessible, opening the door for more academic labs and under-resourced teams to not merely consume but actively contribute to the development of frontier weather models.