When Hurricane Lee was far out in the Atlantic, forecasters faced a familiar challenge: small changes in the storm’s path could determine whether it brought rain to Scotland or serious trouble to the US Northeast. This time, alongside conventional physics-based forecasts, meteorologists had a newer tool to consider: AI weather models.
Lee offered an early comparison
Models developed by Nvidia, Huawei, and Google’s AI unit DeepMind all projected that Lee would make land somewhere between Rhode Island and Nova Scotia. Their outlook generally agreed with the official forecast based on atmospheric physics. Lee eventually reached Nova Scotia on September 16, at the edge of the range predicted by the AI models.
That result encouraged some forecasters, including Mark DeMaria, an atmospheric scientist at Colorado State University who recently retired from leading a division of the US National Hurricane Center. He began a project with the US National Oceanographic and Atmospheric Administration to assess Nvidia’s FourCastNet model against real-time storm data. At first, he was skeptical. The early results changed his view.
The season’s evidence is still limited. Just over halfway through the Atlantic storm season, there had been 16 named storms, but researchers said it was too early to draw final conclusions. So far, AI forecasts have performed comparably to conventional ones, and have sometimes done better at tracking tropical storms.
Learning weather from patterns
Conventional weather models use equations that describe the atmosphere. They take in observations such as temperature, wind, and humidity, then calculate what may happen next. Their accuracy has improved over time as scientific understanding and the volume of observations have grown.
AI weather models take a different route. They learn patterns from decades of atmospheric data in the ECMWF’s ERA5 dataset instead of starting with built-in physics rules. The approach resembles text-generation systems: after training on many examples, a model can reproduce patterns without being explicitly taught each underlying rule.
That does not mean weather prediction becomes straightforward. In the 1960s, meteorologist and mathematician Edward Lorenz showed how small uncertainties in weather data could lead to very different forecasts. He estimated that the atmosphere could be predicted at most two weeks ahead. The field has yet to reach that theoretical limit in practice.
Hurricanes make the AI training challenge sharper. They are only a small part of the available weather data, so a model has relatively few examples of these storms to learn from. The successful track forecasts suggest that AI has picked up important atmospheric patterns, but they do not establish that it will handle every storm well.
Speed helps, but important gaps remain
The clearest advantage so far is speed. A traditional forecast can take hours of supercomputing time, while an already trained AI model can produce a prediction in minutes on a laptop. DeMaria said FourCastNet can run in 40 seconds on an old graphics card.
That speed could matter when forecasters need to compare many possible outcomes. An ensemble forecast presents several scenarios and their likelihoods. For tropical storms, these projections are often called spaghetti models because their possible tracks appear as a tangle of lines. Conventional systems can take hours to calculate each added scenario; AI may make larger groups of projections practical.
But speed does not make the projections complete. Machine-learning models tend to favor common patterns, which can lead them to understate unusual events such as extreme heat waves or tropical storms. They also are not designed to estimate rainfall, which occurs at a finer scale than the global data used to train these models.
Rain and extreme events remain especially challenging. DeepMind’s localized radar-based NowCasting method can predict precipitation, but combining it with global AI weather forecasts is difficult. More detailed data in a future version of the ECMWF dataset may help, and researchers are exploring ways to make models more responsive to rare events.
Forecasters still need to judge uncertainty
Forecasts need to show not just possible weather, but how much confidence to place in each outcome. Ensemble forecasts can account for uncertainty in the starting observations and uncertainty in the model. AI systems can currently help explore variations in the observations, but they do not provide the same assessment of uncertainty in the model itself.
That limitation is tied to the “black box” problem: it can be hard to explain how a machine-learning system arrived at a prediction. Huawei researcher Lingxi Xie said meteorologists most often ask for explanations of AI forecasts, and that the systems cannot yet provide a satisfying answer.
AI forecasts also depend on strong observations in the first place. Satellites, buoys, planes, and sensors supply data that organizations such as NOAA and the ECMWF process into machine-readable datasets. Faster models could eventually make accurate forecasts more widely available, but access to the data that powers them remains essential.
For now, AI is a promising addition to hurricane forecasting rather than a complete replacement for established methods. Its fast projections can help forecasters explore possible storm tracks, while conventional models and human judgment remain important for interpreting uncertainty and assessing the risks AI handles poorly.