AI Predicted a Category 5 Hurricane a Week Out, Using Statistics, Not Physics
Classical numerical forecasting solves fluid equations; AI learns statistical patterns from historical weather. Hurricane Melissa was locked in as a Category 5 while still a tropical depression, at roughly 80% confidence — earlier and more decisively than conventional models.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The model called Category 5 before the storm existed
Hurricane Melissa hit Jamaica in October 2025. At a stage when it was not even a tropical storm — possibly just a tropical depression — DeepMind's model began showing it would become a Category 5 hurricane, with confidence rising over the course of a week from lower levels to about 80%. Peter said that as far as he knew, other models were also predicting intensification at the time, but none had that kind of confidence, that specific track, and that specific intensity. The National Hurricane Center officially determined on Saturday that it would become a Category 5 and strike Jamaica, when it was not even a Category 1. It was the first Category 5 forecast the hurricane center had ever issued from the lowest intensity, and they later said the call was "heavily influenced" by the confidence of DeepMind's model.
— Peter BattagliaForecasting is hard because you cannot observe the butterfly
Peter traces the difficulty of forecasting to a fundamental limit: small things can have large effects, and we cannot observe every butterfly. So there is a fundamental limit in observation, and that limit translates into a fundamental limit in forecasting. But he adds a key turn: that does not mean we cannot detect subtle patterns from all the historical evidence that were previously overlooked. These important features that drive weather, though unobservable, leave "little clues and breadcrumbs" that statistical learning can pick up. He uses the analogy of dropping a small pebble into a pond: if you can catch that little splash and understand it means concentric ripples will spread out, you can use it.
— Peter BattagliaAn AI model can see the whole hurricane at once
GraphCast was one of the first models in the third phase to simulate global weather as a whole out to 10 days. Its approach: take the complete state of global weather, run the same learned local function everywhere (which roughly corresponds to "physics is the same everywhere"), process these into a larger representation covering the globe, then predict back down to the local. Classical numerical forecasting does not do this; it only makes predictions at very local scales. Peter says this may be why AI models are so effective: a large hurricane can span hundreds of kilometres, and structure on the west side can tell you a lot about the east side. If the model can put the entire hurricane into one representation, it can better constrain what happens locally next.
— Peter BattagliaProbabilistic forecasting works by stuffing butterflies into the model
A probabilistic model makes many guesses rather than just one. DeepMind has two approaches. One is diffusion models: start from a large number of different noise images, refine them into something that looks like weather, and each time you land on a slightly different scenario — that is where the diversity comes from. The other is functional generative networks, which they invented: inject different input scenarios and slightly alter the neural network's weight parameters, producing diverse outputs. Peter's metaphor is "what would the world look like if there were a butterfly here and a butterfly there" — run it hundreds or thousands of times and you can see what the future most likely looks like. When all scenarios give the same prediction, confidence is high; when they spread out like spaghetti, the model is saying "we cannot tell what will happen" — and Peter thinks a model must admit that. It is a kind of "humility."
— Peter BattagliaWeatherNext 3 compresses the whole weather tree into one model
Peter uses the metaphor of a "weather tree": the roots are data, the junction of root and trunk is the estimate of the current weather state, the trunk is the operational model that predicts global weather, the branches are regional models for specific uses like energy and extreme events, and the leaves are applications — you pull out your phone to know the temperature right now. Traditionally this value chain is done in at least three or four stages. WeatherNext 3's change: it does not just take the weather state estimate given by a meteorological agency, it eats raw satellite imagery directly; and it does not just predict a global weather estimate, it directly predicts observed values at high-quality weather stations like airports. That is, one model eats from the roots all the way to the leaves. Peter says many groups have been thinking about this idea, but WeatherNext 3 may be the first to significantly beat all previous AI and conventional models on accuracy.
— Peter BattagliaA forecast every hour, because the satellites change every hour
Several concrete changes in WeatherNext 3: higher resolution, no longer blurry or pixelated images, and it can make predictions at different resolutions; forecast frequency goes from once every 6 hours to once every 1 hour, because the satellite imagery it consumes is itself updated hourly; the temporal resolution of the forecast itself is also higher — the model natively predicts a series of weather states within a 6-hour window, such as what the weather is at 1, 2, and 3 o'clock, rather than only saying what it is at 6 and 12. Peter says much of this design came from users telling them what they want in a next-generation weather model — renewable energy forecasting and electricity load forecasting, for example.
— Peter BattagliaThe AI model's failure mode is regressing to average weather
Hannah asks: using the current time step to predict the next, do small errors accumulate and amplify? Peter says this is more a problem of conventional methods — there is nothing in the equations that pulls a drifting solution back to the most reasonable state; once you enter a rare state, the equations will not say "you must behave like normal weather." AI models fail the opposite way: when they start to go wrong, they tend to predict average weather, i.e. "regression to the mean." And predicting average weather is actually a fairly good way to forecast, and the oldest way — farmers planting in spring rely on seasonal patterns. So AI models are less likely to "blow up" the way conventional numerical methods do. But Peter also concedes that why statistical learning can predict the strongest hurricane on record "probably still needs to be explored better."
— Peter BattagliaMost weather observations are not used at all
Peter says weather itself affects roughly a third of the economy, but better forecasts do not necessarily help that third — it is just that weather touches energy, agriculture, insurance, transport, and disasters. He thinks the room for the next step is enormous: industries like energy are themselves collecting large amounts of weather observations that could be folded into models, and they have not done that yet; many weather phenomena occur at scales smaller than a kilometre; and "most weather observations are not actually used by meteorological agencies." His judgment: we have only proved we can make very accurate forecasts; now the real work begins. He also says he will move away from "approximately solving equations," but not away from physics — the data itself was made by physics, it is just a different way of approximating physics.
— Peter BattagliaIn their own words · checked verbatim
we started to see the model becoming quite confident that it was going to turn into a category 5 hurricane, which is the strongest category of hurricane.
Peter Battaglia2:10
this was the lowest intensity storm that the National Hurricane center had ever forecast to become category 5. They they had never made a category 5 forecast from such a low intensity.
Peter Battaglia5:10
we can't observe every butterfly. So there's a fundamental limit to what we can observe and that then translates to a fundamental limit in what we can predict.
Peter Battaglia12:11
if they're spaghetti plots and they're all over the place then the model is saying it's very difficult for us to understand what could happen and that's a very important thing because it's a fact of how we forecast weather. We cannot know certain things.
Peter Battaglia25:17
With AI models, their failure mode tends to be when they start to make errors, they tend to more just predict average weather. Like predicting the average weather is actually a pretty good way to predict weather. And it's actually the oldest way to predict weather.
Peter Battaglia36:22
while the model might not have seen this specific instance of this storm at this location and this trajectory, it has seen intense weather in other parts of the globe
Peter Battaglia37:22
most weather observations are actually not even used by weather agencies. So, there's I actually think there's a ton of data. There's a ton of problems that haven't even been touched yet.
Peter Battaglia39:22
I think we will move away from approximating the solutions to equations, but we're not moving away from physics. The data was also created by physics.
Peter Battaglia40:23
Figures
| Hurricane Melissa maximum sustained winds at landfall | 190 mph | 5:10 |
| Hurricane Melissa death toll | 95 people | 5:10 |
| DeepMind model confidence in its Melissa forecast | about 80% | 5:10 |
| Accuracy gain of AI forecasting over conventional forecasting | about one extra day (a two-day-accuracy forecast is reached at three days) | 9:11 |
| WeatherNext 3 forecast update frequency | every 1 hour (previously every 6 hours) | 29:18 |
| Share of the economy affected by weather | about one third | 38:22 |
| GraphCast forecast horizon | 10 days | 20:14 |
Glossary
- numerical weather prediction
- The classical forecasting method that uses supercomputers to approximately solve the equations of fluid motion and extrapolate step by step.
- GraphCast
- An early DeepMind global AI weather model that predicts global weather 10 days out in one pass.
- GenCast
- DeepMind's probabilistic weather model, which predicts multiple possible scenarios rather than a single average weather state.
- functional generative networks
- A probabilistic forecasting technique developed by DeepMind that generates diverse scenarios by slightly adjusting network weights.
- spaghetti plots
- Plots that draw the tracks of multiple probabilistic scenarios together; the more spread out they are, the greater the uncertainty.
- load forecasting
- Predicting electricity demand on the grid, which is mainly driven by temperature.
How to listen
Engineers and researchers working on AI for Science, weather, or energy forecasting; investors watching AI land in real-world settings.
0:38–11:50, the Hurricane Melissa narrative stretch, can be fast-forwarded. The mechanisms and numbers are concentrated in the second half.