On July 9, 2026, Microsoft released "Aurora 1.5," an AI weather model, along with a pre-print technical paper. The biggest change is the shift from a model that produces a single forecast to one that runs multiple possible futures to show probability and uncertainty. In a retrospective evaluation using tropical cyclones from 2024–2025, the median track derived from 32 forecast runs reduced errors by 13–34% compared to the previous Aurora. This adds a new tool to typhoon and hurricane track forecasting.
However, it would be inaccurate to summarize this simply as "hurricane forecasting improved by 34%." What the paper measured was the track position—where the center of the storm passes. It did not evaluate intensity, precipitation amounts, wind field extent, or storm surge. Furthermore, the comparison target, the European Centre for Medium-Range Weather Forecasts' (ECMWF) ENS, is a conventional physics-based system. How Aurora 1.5 compares against ECMWF's currently operational AI ensemble, "AIFS ENS v2," remains unknown.
From 4 Variables to 26, Expanded to Hourly Intervals
The original Aurora directly handled only 4 surface-level variables: 2m temperature, 10m wind, and mean sea-level pressure. Aurora 1.5 adds 22 variables, expanding the total to 26. Beyond cloud cover and precipitation, it now forecasts surface solar radiation and 100m wind. Soil moisture and snow cover are also included.
This expansion means more than just refining a typical weather map. 100m wind relates to wind power generation, while cloud cover and solar radiation affect solar power output. The timing of precipitation onset and gusts influence decisions in aviation and logistics. Microsoft Weather cites applications in energy, agriculture, and transportation. The Swiss utility company BKW is already combining Aurora 1.5 with its existing operational models.
The forecast interval has also become finer, shifting from the conventional 6-hour focus to a native 1-hour resolution. During training, the research team varied forecast intervals from 0 to 12 hours, enabling the same model to generate hourly states. This design is aimed at use cases where a difference of a few hours—such as the onset of precipitation or landfall timing—can change decision-making.
The code is published on GitHub under the MIT License, and model checkpoints are also distributed. That said, Microsoft itself positions this as being for research evaluation purposes. Automated decision-making for disaster response or critical infrastructure without expert verification is not among the intended use cases.
Learning a Distribution That Includes How Forecasts Can Diverge, Rather Than a Single Prediction
Weather forecasting faces the fundamental problem that initial conditions can never be measured perfectly. Even tiny observational errors cause forecasts to diverge as time progresses. ECMWF's physics-based ENS generates 50 forecast members by slightly perturbing initial conditions and the model's physical processes, plus a control forecast, showing a range of possibilities up to 15 days ahead at approximately 9km resolution.
Aurora 1.5 ENS also produces multiple forecasts, but the method of generating divergence differs. The research team injects Gaussian noise into the normalization layers within the neural network, generating a different future each time inference is run. The training objective adopts CRPS (Continuous Ranked Probability Score). Rather than selecting a single forecast closest to the correct answer, this metric highly rates forecast distributions that appropriately envelop the observed outcome without spreading unnecessarily wide.
A key advantage is the ability to reuse the existing foundation model. The deterministic and probabilistic versions were built through staged incremental training. Using up to 32 NVIDIA A100 80GB GPUs, the main variable-expansion stage took 2 weeks, and the main probabilistic stage took 10.5 days. Since the paper does not disclose the time or cost required for a single forecast, it's impossible to judge how much lower operational costs might be compared to conventional systems.
What a Reduction of Up to 34% Actually Means: "Track Position" Error
The tropical cyclone evaluation used cases that occurred globally in 2024–2025, along with the IBTrACS track records finalized after observation. Forecasts were issued 4 times daily, verified against 1–5 day forecasts at 6-hour intervals. In this test, no variation was given to initial conditions—all 32 runs were launched from the same ECMWF control initial condition—so the distribution primarily represents internal model uncertainty. The deterministic version of Aurora 1.5 reduced track errors by 9–24% compared to the previous Aurora. Taking the median track from the 32 forecast runs widened the improvement to 13–34%. The aggregate figure cited in the paper's abstract is a 16% reduction.
For Hurricane Helene, in a forecast issued at 00:00 UTC on September 24, 2024, the average track error was 110.6km for the U.S. National Hurricane Center's (NHC) official forecast, 58.4km for Aurora 1.5, and 26.3km for the Aurora 1.5 ENS average. It also captured the landfall location, with 24 of the 32 runs enveloping the actual track. On the other hand, it predicted movement approximately 6 hours faster than actual.
The numbers from this single Helene case should not be generalized into a broad win rate against the NHC. The NHC's official track forecast integrates multiple models along with forecaster judgment. In 2024, the Atlantic basin saw record-best average track accuracy across all forecast periods from 12 to 120 hours, surpassing individual models at the 96-hour and 120-hour marks. The value Aurora 1.5 adds lies not in replacing the official forecast, but in providing forecasters with one more independent distribution for comparison.
Don't Confuse "Beating ECMWF in 88.9% of Cases" With the Current State of AI Competition
Microsoft's paper reports that Aurora 1.5 ENS achieved better CRPS scores than ECMWF's physics-based ENS in 88.9% of the evaluated variable and forecast-time combinations. The targets were upper-atmosphere geopotential height, temperature, specific humidity, and 5 surface-level variables, measured across 1–10 day forecasts. The comparison used 103 forecasts initialized on Mondays and Thursdays in 2024, comparing both models with 50 members each.
This 88.9% figure does not represent a win rate across all weather conditions, regions, or disaster types. On the Aurora side, perturbed initial conditions generated by ECMWF ENS were used as inputs for the 0-hour and 6-hour timesteps. In the first 18 hours, it underperformed ECMWF ENS for some variables, and weaknesses remain at the 10-day mark for 2m temperature and 200hPa geopotential height. While widening the forecast spread improved CRPS, the research team also acknowledges a tendency toward overly broad spreads.
Competitors haven't been standing still either. On May 12, 2026, ECMWF put its AI ensemble, AIFS ENS v2, into operational use. It computes forecasts up to 15 days ahead at approximately 31km resolution, and for tropical cyclones, offers 52 members including AIFS Single. Aurora 1.5's paper compared against the physics-based ENS but did not include a cross-comparison with AIFS ENS v2 or other probabilistic AI models.
Aurora 1.5 stands as a powerful implementation moving AI weather forecasting from producing "a single most-likely outcome" toward representing "a range of possible outcomes." Whether this publicly released model becomes practically viable depends on whether, during the next typhoon season, it can be validated in real time not just for track but also for intensity and precipitation—and how much it can improve forecasters' judgment within an operational system that combines existing physics-based models with AIFS ENS v2.