Microsoft has released Aurora 1.5, a major update to its open-source Earth system foundation model that expands its coverage from four to 26 weather variables and adds one of the most requested capabilities in operational forecasting: ensemble predictions.
The new variables cover surface conditions, pressure-level fields, wind, temperature, humidity, precipitation, and radiation, making the model relevant to sectors from energy and agriculture to transport and climate risk planning. The shift to hourly temporal resolution enables finer-grained operational guidance for events like precipitation onset or tropical cyclone landfall.
Ensemble Forecasting
The ensemble version introduces stochastic perturbations to represent model uncertainty, generating multiple forecast members to estimate the range of possible outcomes. This matters because a single best-guess forecast is often less useful than knowing the probability distribution, especially for high-stakes decisions in power systems, extreme-weather planning, and climate risk.
Microsoft says Aurora 1.5's probabilistic forecasts outperform the state-of-the-art ECMWF dynamical ensemble on 88.9% of evaluated variable-and-lead-time targets. In tropical cyclone testing across all 2024-2025 storms, the model reduced track errors substantially, with the ensemble median showing roughly one-third lower error by day five compared to the original Aurora.
The ensemble capability was built through multi-stage fine-tuning on top of the base Aurora model, with a final round of auto-regressive fine-tuning on ECMWF HRES analysis data from 2018 to 2023 to improve rollout stability.
From Research to Operations
Aurora began in Microsoft Research AI for Science and is now being built on for operational use by Microsoft Weather. The model is released as open source on GitHub with checkpoints on Hugging Face, while Microsoft Weather provides the data infrastructure, managed access, and operational assurance that enterprises need.
This dual-track approach, open for researchers and supported for production users, reflects a broader question in AI for science: how to bridge the gap between a promising research model and something organizations can bet decisions on.
Microsoft Weather says it plans to extend this work with additional fit-for-purpose AI weather models designed for enterprise scenarios where forecast quality, speed, uncertainty quantification, and operational decision support matter most. The team has been applying AI to operational forecasting for over seven years and was ranked the world's most accurate global forecast provider by an independent third party for three consecutive years from 2022 to 2024.
For researchers and agencies, the open release means the ability to evaluate, adapt, and extend the model. For enterprises, the Microsoft Weather path offers infrastructure and support. The underlying message is that foundation models should complement, not replace, physics-based models and domain expertise, but only if they are deployed with careful evaluation and transparency.