From Research to Operational Weather Intelligence
Microsoft has released Aurora 1.5, a major update to its open-source Earth-system foundation model that adds 22 new weather variables, hourly temporal resolution, and probabilistic ensemble forecasting. The model, developed by Microsoft Weather as an extension of the original Microsoft Research AI for Science release, is available on GitHub and Hugging Face for researchers and developers to evaluate and extend.
The update reflects a deliberate strategy: keep the research foundation open while building enterprise-grade operational capabilities on top. As climate and weather risks intensify globally, the ability to generate fast, accurate, and uncertainty-aware forecasts is becoming critical for energy grids, agriculture, transport, and disaster preparedness.
What's New in 1.5
The breadth expansion adds 22 variables to Aurora's original four, covering surface conditions, pressure-level fields, wind, temperature, humidity, precipitation, and radiation. This makes the model relevant for sectors that depend on integrated Earth-system signals rather than just temperature and precipitation.
The shift to hourly temporal resolution enables fine-grained operational guidance for events like precipitation onset, tropical cyclone landfall, and trade wind shifts. Perhaps most significantly, Aurora 1.5 adds ensemble forecasting, which runs multiple simulations to estimate the range and likelihood of possible outcomes. This is essential for decision-making under uncertainty, where the distribution of futures matters as much as the single best estimate.
Outperforming the State of the Art
In head-to-head evaluation against the ECMWF dynamical ensemble, Aurora 1.5 outperformed on 88.9% of evaluated variable-and-lead-time targets across upper-air geopotential, temperature, humidity, and five surface variables. In tropical cyclone track forecasting, the model reduced track errors substantially, with the ensemble median showing roughly one-third lower error than the original Aurora by day 5. For Hurricane Helene, the ensemble effectively captured the uncertainty in the storm's progression, enveloping the verified track.
The ensemble capability was developed through multi-stage fine-tuning on ECMWF HRES analysis data from 2018 to 2023, with controlled perturbations introduced into the model's latent conditioning pathway.
Open Research, Managed Operations
Aurora 1.5's open-source release is intended to let researchers, agencies, and companies evaluate and extend the model. Microsoft Weather is building managed services and operational deployment paths for organizations that need additional data, infrastructure, and decision-support capabilities. This dual-track approach, open research plus managed operations, is designed to bridge the gap between scientific advance and practical application.
Endesa, a major European utility, is already using Aurora 1.5 alongside existing operational Microsoft Weather models to support energy operations where weather-dependent generation and infrastructure planning converge. The UK Met Office is exploring how foundation models can complement established physics-based systems across weather and climate time scales.
Beyond Forecasting: Climate Applications
The model's applications extend beyond traditional weather prediction. Terradot, a portfolio company of the Microsoft Climate Innovation Fund, is using Aurora to estimate and optimize carbon dioxide removal from enhanced rock weathering under real field conditions. The company reports that building on Aurora is significantly advancing its R&D timelines toward gigaton-scale carbon removal.
Context and Competitive Landscape
Aurora 1.5 enters a field that includes other AI weather models such as Google's GraphCast, NVIDIA's FourCastNet, and various ECMWF experimental systems. Microsoft's differentiation appears to be the combination of open model weights, enterprise integration through Azure, and a explicit connection to operational Microsoft Weather services that already reach over a billion devices globally.
The model's performance gains, particularly in ensemble forecasting and tropical cyclone tracking, suggest that AI weather models are approaching or exceeding the skill of traditional numerical weather prediction in specific domains. The critical question for the field is whether these gains hold across all weather regimes and how quickly operational meteorological services can integrate AI forecasts into their existing warning and decision frameworks without compromising the trust built over decades of physics-based modeling.