Aurora 1.5: Microsoft opens ensemble weather model, 22 vars

Microsoft's Aurora 1.5 is a major extension of the open Aurora Earth-system foundation model, released as open source on GitHub with model checkpoints on Hugging Face. The update adds 22 weather variables, moves to hourly temporal resolution, and introduces probabilistic ensemble forecasting through stochastic perturbations and multi-stage fine-tuning.

What is Aurora 1.5?

Aurora 1.5 is an expanded version of the Aurora foundation model that adds 22 new weather variables to the original model's 4, plus hourly forecasts and an ensemble capability. The original Aurora was introduced in 2024 by Microsoft Research AI for Science and published in Nature in 2025. Aurora 1.5 is offered openly so researchers and developers can evaluate and extend it, while Microsoft Weather connects the model to data, infrastructure, managed access, and operational use for enterprises.

Beyond the variable and temporal updates, Aurora 1.5 was fine-tuned with an auto-regressive pass on ECMWF High Resolution (HRES) analysis data from 2018 to 2023 to improve rollout behavior and stability. Microsoft presents Aurora 1.5 as both an open research foundation and a platform that can be integrated into operational workflows through Microsoft Foundry, Planetary Computer Pro, Agent skills and Azure services.

How does Aurora 1.5's ensemble perform?

Aurora 1.5's probabilistic ensemble forecasts outperform the ECMWF dynamical ensemble on 88.9% of the evaluated variable-and-lead-time targets, and ensemble-based forecasts reduced tropical cyclone track errors substantially compared with the original model. The ensemble implementation introduced controlled perturbations into the model's latent conditioning pathway, then optimized for probabilistic forecast quality across multiple forecast members.

Microsoft highlights a diagnostic comparison showing that across upper-air geopotential, temperature, humidity and five surface variables, Aurora 1.5 outperforms ECMWF ENS on 88.9% of evaluated targets. For tropical cyclones tested during 2024 to 2025, Aurora 1.5 reduced track error relative to the original Aurora, with the ensemble median reaching roughly one-third lower error by day 5. The release includes an example showing the ensemble around Hurricane Helene starting at 0 UTC on September 24, 2024, where multiple plausible tracks envelop the verified track.

The ensemble capability was developed through multi-stage fine-tuning on top of Aurora, adding stochastic perturbations to represent model uncertainty and generating multiple members to estimate forecast spread. This probabilistic output is positioned as important for use cases where the distribution of outcomes matters as much as the best estimate, including power systems, transport, agriculture, and extreme-weather planning.

Why it matters

Aurora 1.5 bridges open scientific research and enterprise-grade weather intelligence by combining an open foundation model with product engineering and cloud infrastructure. Microsoft says the release enables researchers, agencies, companies, and civil society to evaluate and extend the model, while Microsoft Weather supplies managed services and operational pathways where additional data and assurance are required. Sridhar Iyer, Corporate Vice President, Microsoft AI, framed the release as "a meaningful step toward making weather foundation models more open, useful, and practical."

The update expands the model's relevance beyond short-term forecasting to sectors that need integrated Earth-system signals, and Microsoft points to partnerships and applications beyond forecasting, including work with Terradot on carbon removal R&D and collaborations with the UK Met Office to explore how foundation models can work alongside physics-based systems.

Microsoft Weather also notes operational scale: the team reaches more than a billion devices across 180 countries, has applied AI to operational forecasting for more than 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.

What to watch

Watch for the next set of fit-for-purpose AI weather models Microsoft plans to release in the coming months, and for further evaluations comparing Aurora-derived forecasts with established dynamical systems in operational settings. Adoption signals to track include integrations into enterprise workflows, partner deployments such as BKW's energy operations tests, and additional benchmarks against ECMWF and other operational ensembles.