Microsoft just upgraded its open-source AI weather model, and the latest version promises to change how we plan our days and respond to severe weather. Aurora 1.5, released by Microsoft Research, now delivers hourly forecasts powered by ensemble simulations, covering 22 key atmospheric and oceanic variables. It’s a major leap from its predecessor, and it’s available to anyone.
A Deeper Look at Aurora 1.5’s New Capabilities
Despite its quiet launch, Aurora 1.5 packs three critical enhancements: higher temporal resolution, more variables, and probabilistic guidance. Where earlier versions generated forecasts at coarser time steps, Aurora 1.5 steps down to hourly intervals. This means you could get a more precise picture of when a storm might intensify or when temperatures will peak.
The model’s variable count jumps to 22, now including not just temperature and wind but also humidity, precipitation, cloud cover, and ocean surface conditions. This breadth makes it a true Earth-system model, capable of simulating interactions between the atmosphere, land, and ocean. Among the newly added variables are specific humidity at multiple pressure levels, surface pressure, geopotential height, and sea-surface temperature—essential ingredients for understanding weather dynamics.
Most importantly, Aurora 1.5 introduces probabilistic ensemble forecasting. Instead of a single deterministic prediction, it runs multiple simulations with slightly different initial conditions to produce a range of possible outcomes and their likelihoods. So rather than just saying “it will rain at 3 p.m.,” the model can indicate a 70% chance of rain between 2 and 4 p.m., giving planners critical uncertainty information. This is the first time such capability has been openly available in an AI-based Earth-system model.
The Technical Underpinnings
Aurora is a foundation model—a large neural network pretrained on decades of historical weather data and reanalysis products like ERA5. It learns the underlying dynamics of the atmosphere, oceans, and land surface, encoding them into billions of parameters. This pretraining allows it to generate forecasts in seconds once given current observational data, bypassing the need to solve complex physical equations from scratch. The 1.5 version builds on this base with expanded output heads for the additional variables and a new mechanism to generate ensemble spreads efficiently.
Why Hourly, Ensemble Forecasts Matter
Traditional numerical weather prediction (NWP) relies on solving fluid dynamics equations on supercomputers, a process that can take hours and consume megawatt-hours of energy. AI models like Aurora learn from historical data to mimic this process in seconds. With hourly updates, Aurora 1.5 can be rerun frequently as new observations arrive, keeping forecasts fresh. This is a game-changer for rapidly evolving weather events like thunderstorms or flash floods.
Ensembles have long been the gold standard in professional meteorology because they quantify forecast confidence. For example, hurricane track forecasts often show a “cone of uncertainty.” Aurora’s ensemble approach brings that same probabilistic thinking to routine weather, helping agriculture, energy traders, and event planners make better decisions. An energy utility, for instance, could use probabilistic wind forecasts to decide how much reserve capacity to keep online, potentially saving millions.
For everyday users, this might mean a weather app on your phone that warns you of a sudden downpour with enough lead time to grab an umbrella—and tells you how sure it is about that prediction. No more wondering if that 30% chance of rain really means bring an umbrella.
Who Gains from Aurora 1.5
Everyday Users and Small Businesses
If you check the weather before a picnic or a commute, Aurora 1.5’s hourly ensembles could make those quick glances far more reliable. Weather services and app developers can integrate the model to provide free, high-resolution forecasts. And because it’s open-access, small businesses can build custom alert systems without paying for proprietary data. A landscaping company could receive tailored notifications about frost probability, while a food truck operator might get precise rain timing.
Researchers and Universities
The open-source nature of Aurora 1.5 is a boon for climate science and meteorology programs. Students and researchers can experiment with the model, fine-tune it for regional climates, or study how AI captures atmospheric dynamics. Because the model is a foundation model, it can be adapted with relatively little additional data to specific domains—such as predicting air quality in a city or optimizing wind farm operations. Microsoft Research has a history of releasing models with permissive licenses, and Aurora appears to follow suit, encouraging academic exploration.
Emergency Managers and Humanitarian Groups
Hurricane tracking is one of the most high-stakes applications. Faster, more frequent forecasts with uncertainty ranges can improve early warnings for tropical cyclones, floods, and heatwaves. Organizations like the Red Cross or FEMA could use Aurora to prioritize resources ahead of disasters. In regions with sparse weather infrastructure, such as sub-Saharan Africa or small island nations, an open model running on modest hardware could provide actionable warnings that were previously unavailable.
The Equity Angle
Traditional NWP models require supercomputers that only national weather services in wealthy countries can afford. AI models like Aurora run on consumer-level GPUs, making advanced weather prediction accessible to developing nations, independent researchers, and even citizen scientists. By open-sourcing Aurora 1.5, Microsoft is lowering the barrier to entry for all, potentially catalyzing a wave of localized weather services worldwide.
How We Got Here: The Evolution of AI Weather Models
The journey toward AI-driven weather forecasting accelerated in 2022 with Nvidia’s FourCastNet, which showed that a neural network could rival traditional models on a desktop GPU. In 2023, the field exploded: DeepMind’s GraphCast delivered medium-range forecasts that beat the European Centre for Medium-Range Weather Forecasts (ECMWF) in many benchmarks; Huawei’s Pangu-Weather proved AI could handle tropical cyclone tracks; and Microsoft Research introduced Aurora, framing it as an Earth-system foundation model rather than a pure weather predictor.
Aurora stood out by training on a broader set of data, including ocean and chemistry variables, and by being open-access from day one. The original version could produce 10-day forecasts at 6-hour time steps for a limited set of surface and upper-air variables. Aurora 1.5 expands that vision dramatically, both in temporal granularity and in scope of variables, while adding the ensemble capability that forecasters crave.
This evolution mirrors a broader shift in scientific computing. Foundation models pretrained on massive datasets are proving more adaptable and efficient than single-purpose simulators. With Aurora 1.5, Microsoft is signaling that Earth system science will be one of the primary beneficiaries.
What to Do Now: Access Aurora 1.5
If you want to try Aurora 1.5 yourself, Microsoft has made it available under an open-access license. Here’s how to get started:
-
Visit the Aurora project page
Head to Microsoft Research’s official Aurora repository (link in references). You’ll find documentation, a model card describing the architecture and intended uses, and Jupyter notebooks that walk through a demo. -
Set up your environment
While you can run inference on a decent GPU with 16 GB of memory, for full-scale hourly ensembles you may want cloud resources. Microsoft Azure provides ready-made templates, or you can use any Python environment with PyTorch 2.0+. A pip install from the repository typically covers dependencies. -
Fetch initial conditions
The model needs current global weather data as input. The repository includes scripts to pull data from public sources like NOAA’s Global Forecast System (GFS) or ECMWF’s operational analyses in real time. You’ll need an internet connection and about 10 GB of storage for a full starting state. -
Run forecasts and visualize
The code outputs standard netCDF files; you can visualize them with Python libraries like Matplotlib or integrate them into a web dashboard. A single ensemble run (say, 30 members) takes about a minute on an A100 GPU. The output can be served as an API if you want to build an app. -
Join the community
Microsoft Research maintains a discussion board and a Slack channel where researchers and developers share ideas, report issues, and showcase applications. You’ll find examples of regional downscaling, machine learning for impact assessment, and even attempts to predict atmospheric rivers.
For non-developers, the impact will trickle down through weather apps and services that adopt the model. You likely won’t need to do anything except maybe choose an app that uses Aurora’s forecasts once they become available.
The Outlook: Smarter, Faster, Free
Aurora 1.5 is a milestone, but Microsoft’s research team isn’t stopping here. Expect even higher spatial resolution—perhaps down to 0.1 degrees—longer lead times, and tighter integration with Microsoft’s cloud and AI platform. The company has hinted at applying similar foundation models to other Earth system challenges like carbon monitoring, wildfire spread, and ocean health.
In the near term, the open availability of hourly ensemble forecasts could spark a wave of innovation from the global developer community. Imagine hyperlocal weather apps for farmers in sub-Saharan Africa, shipping companies optimizing fuel consumption based on highly probable wind patterns, or insurance firms pricing risk more accurately with bottom-up hazard models. That’s the promise Aurora 1.5 brings—and it’s now anyone’s to build on.