AI
Microsoft Releases Aurora 1.5, Expanding Open-Source Weather AI with 22 New Variables and Ensemble Forecasting
An upgraded version of Microsoft's Earth-system foundation model now offers broader atmospheric data, hourly forecasts, and probabilistic ensemble predictions for energy, agriculture, and climate applications.
Microsoft Research Blog
Microsoft has unveiled Aurora 1.5, a significantly enhanced iteration of its open-source Earth-system foundation model designed to support weather forecasting and climate intelligence across multiple sectors. The update dramatically expands the model's analytical capabilities while maintaining its status as freely available research infrastructure for global use.
Expanded Scope and Enhanced Precision#
The refresh brings substantial new depth to Aurora's forecasting ability. Where the original model tracked four core weather variables, Aurora 1.5 now monitors 26 distinct atmospheric and Earth-system indicators—including surface temperature, humidity, precipitation, wind patterns, and solar radiation measures. This broader instrumentation makes the model significantly more useful for industries sensitive to integrated atmospheric signals, such as renewable energy operators, agricultural planners, and logistics networks.
The upgrade also doubles temporal granularity, shifting from 6-hour to hourly resolution. This finer detail proves critical for time-sensitive operational decisions: detecting the precise moment precipitation begins, anticipating tropical cyclone landfall timing, or managing real-time trading exposure to weather swings.
Ensemble Forecasting Captures Uncertainty#
One of the most significant additions—ensemble probabilistic forecasting—directly addresses a longstanding user request. Rather than producing a single deterministic forecast, Aurora 1.5 now generates multiple simulation runs that collectively illustrate the range and likelihood of possible atmospheric outcomes. This approach accounts for the inherent sensitivity of weather systems to initial conditions and modeling assumptions.
The ensemble capability was developed through iterative refinement, beginning with expansion of the variable set and temporal resolution, then introducing controlled perturbations into the model's latent pathways, and finally optimizing probabilistic accuracy against historical European Centre for Medium-Range Weather Forecasts (ECMWF) data from 2018–2023.
Performance testing demonstrates tangible gains: Aurora 1.5's ensemble predictions outperform the ECMWF operational ensemble on roughly 89% of evaluated metrics across different forecast horizons and weather variables. For tropical cyclone forecasting, the model achieved approximately one-third lower track-error rates compared to the original Aurora, with Hurricane Helene serving as a high-impact real-world validation case.
