Microsoft Releases Aurora 1.5 Weather Foundation Model

Microsoft Research released Aurora 1.5, an updated Earth-system foundation model for weather forecasting, adding 22 weather variables, hourly temporal resolution, and probabilistic ensemble forecasting. The model outperforms ECMWF ensemble forecasts on 88.9% of evaluated targets and reduces tropical-cyclone track error, marking a shift of foundation-model techniques into operational, uncertainty-aware domains like energy, agriculture, and climate-risk planning.

Microsoft Releases Aurora 1.5 Weather Foundation Model Applied ML teams should notice Aurora 1.5 because it turns an AI weather model into a more practical open foundation model for operational forecasting. Microsoft Research says the release adds 22 more weather variables, hourly temporal resolution, and probabilistic ensemble forecasting to Aurora, with code and checkpoints available for researchers and developers. The update matters beyond weather specialists: it is a concrete example of foundation-model methods moving into scientific and infrastructure domains where uncertainty, domain evaluation, and operational trust matter. Microsoft also connects Aurora 1.5 to Microsoft Weather services, Foundry, Planetary Computer Pro, and partner use cases in energy and climate-risk planning. Why it matters Aurora 1.5 is a useful signal for applied ML because it shows foundation-model techniques moving into a domain where accuracy, uncertainty, and operational reliability matter more than chatbot fluency. Microsoft Research says the update extends the Aurora Earth-system foundation model with more variables, hourly forecasts, and probabilistic ensemble forecasting, while keeping the model available for researchers and developers to evaluate and build on. What changed Microsoft Research published Aurora 1.5 on July 9, 2026. The release adds 22 weather variables to the original model, expands temporal resolution to hourly forecasting, and introduces ensemble forecasts that estimate uncertainty across multiple possible futures. Microsoft says the new variable set includes surface, pressure-level, wind, temperature, humidity, precipitation, and radiation fields, which broadens the model's relevance for energy, agriculture, transportation, and resilience planning. The blog also reports that Aurora 1.5 outperforms ECMWF ensemble forecasts on 88.9% of evaluated variable-and-lead-time targets and reduces tropical-cyclone track error versus the original Aurora model in 2024-2025 evaluations. Those figures should still be read as Microsoft-reported evaluation results, but they make the release more substantial than a routine model card update. Practitioner read For data-science teams, the lesson is not simply that weather forecasting has a new model. Aurora 1.5 is an example of a domain foundation model packaged with open access, evaluation claims, cloud infrastructure, and operational pathways. That combination is what applied ML teams increasingly need when moving from research artifacts to decision-support systems. The release also highlights the importance of uncertainty-aware outputs. In areas such as energy demand, storm planning, logistics, and climate risk, a single point forecast is often less useful than a distribution of plausible outcomes. Aurora 1.5's ensemble forecasting makes that uncertainty part of the modeling interface, which is the right direction for high-stakes applied AI systems. Key Points - 1Microsoft Research released Aurora 1.5 with more variables, hourly forecasts, and probabilistic ensemble forecasting for applied weather modeling. - 2The model is positioned for weather, climate, energy, agriculture, transport, and resilience-planning applications that need uncertainty-aware forecasts. - 3Aurora 1.5 shows how open foundation models can move from research into uncertainty-aware operational decision support. Scoring Rationale Aurora 1.5 is notable because it extends an open Earth-system foundation model with practical forecasting features and reported gains against a strong operational baseline. The score stays below 7.5 because it is a domain-specific applied ML release, not a broad general-purpose model or independently validated industry-wide benchmark shift. Sources Public references used for this report. Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. Try 250 free problems /problems