Aurora 1.5: Fine-Tuning a Foundation Model for Medium-Range Ensemble Weather Prediction

Summary

Aurora 1.5: Fine-Tuning a Foundation Model for Medium-Range Ensemble Weather Prediction - Jonathan Weyn , - Zekun Ni , - Amit Misra , - Will Fein , - Haiyu Dong , - Wessel P. Bruinsma , - Richard E. Turner , - Matt Corey , - Kit Thambiratnam , - Kevin White , - Kenji Takeda We present Aurora 1.5, a fine-tuned variant of the Aurora atmospheric foundation model [Bodnar et al., 2025] optimized for skillful medium-range ensemble weather prediction. Building on Aurora’s pretraining across diverse heterogeneous atmospheric data, we introduce a three-stage fine-tuning pipeline that (i) expands the single-level variable set and randomizes lead-time embeddings to enable a native one-hour temporal resolution, (ii) injects Gaussian noise into AdaptiveLayerNorm (AdaLN) modules to generate stochastic forward passes and optimizes a Continuous Ranked Probability Score (CRPS) objective in place of a deterministic loss, and (iii) performs auto-regressive fine-tuning on operational ECMWF analyses with multi-step rollouts. Aurora 1.5 ENS outperforms the ECMWF ENS operational ensemble on 88.9% of upper-air and single-level target variables in the medium range (days 1–10). Reliability diagnostics including rank histograms indicate that Aurora 1.5 ENS achieves this result with a slight over-dispersion, in contrast to the under-dispersion typical of dynamical models. Compared to Aurora, Aurora 1.5 also better predicts extreme events, reducing tropical cyclone track errors by 16% and the mean absolute error on top-5th percentile heat waves by 58%. Aurora 1.5 demonstrates that foundation model fine-is a viable, cost-effective path toward reliable probabilistic AI weather prediction.