How AI is changing weather forecasting

Start with how forecasting works without AI, because the AI story only occupies one part of it. The atmosphere is a fluid, fluids are governed by the Navier-Stokes equations, and those equations are too hard to solve exactly at planetary scale. So the field approximates them numerically and runs the approximation on supercomputers, which is what “numerical weather prediction” means. Battaglia calls the result a triumph: decade on decade the useful horizon stretched out to ten, twelve, fifteen days. But prediction is only the second half of the job. The first half is working out what the weather is right now, by fusing satellites, ground stations, balloons and ships into an estimate of the global state. Because the atmosphere is chaotic, a perfect forecast would require knowing where every butterfly is, so that estimate is always incomplete and forecasts are irreducibly probabilistic. That is why your phone says a chance of rain rather than rain. There is also a third structural feature worth holding onto: the industry is a pipeline. Public bureaus issue coarse global forecasts, and a downstream post-processing layer specializes them for particular places and uses. Machine learning showed up in that post-processing layer first, sharpening the last mile rather than touching the base forecast.

The current wave targets the prediction half instead, and the method is less exotic than the branding implies. It is supervised learning: feed the model a state, train it to produce the next state, then feed its own output back in to roll the forecast forward step by step. What the newer architectures add is reach. A convolutional network learns a local operation and stacks layers until information can cross an image, whereas a transformer, which Battaglia treats as essentially a graph neural network with connections at any distance, links far apart points directly. The pointed detail is what they use that reach for. Language models look backward across a long stretch of text because the next word depends on context from far earlier. Weather does not work that way; it is Markov, meaning the present state in principle determines the next one, so his team deliberately does not let the model look far back in time. They spend the long-range connections on space instead. And that, he suggests, is where the qualitative difference lives: a traditional simulator resolves a hurricane strictly locally out of pressure, temperature, wind and moisture, while a model that can see the whole structure at once appears to treat the hurricane almost like a macroscopic object sliding across the globe. He hedges that twice, with “almost like” and with a flat admission that the field does not understand how the models actually forecast it.

Then the data question, where Kann sets up an intuition the guest partly overturns. Robotics is starved of training data, but weather should be the opposite: a long historical record in which every observation is paired with the measurement that followed. Battaglia agrees the historical record was a windfall, specifically ERA5, a decades-long reconstruction of Earth’s weather built by the European forecasting centre for climate purposes and not for machine learning at all, which happened to be almost perfectly shaped for it. But the record is uneven, because the satellites and stations behind it changed over the decades, so accuracy degrades as you go back. More importantly, you cannot make more of it. Weather has to happen before it can be recorded, and with models stepping in six-hour increments, a day of waiting buys a day of data. Hence the search for stranger inputs: cheap rooftop weather stations, doorbell cameras, car thermometers and rain-sensing wipers, even people posting about the weather, all of which he flags as unproven in quality. His generalization is that in modern AI you are always data poor, and he volunteers that he personally doubted scaling would keep working and was wrong about it.

What this buys is the least settled part of the conversation, because Battaglia is describing possibilities rather than results. Nothing in the episode quantifies how much better an AI forecast is. He is proud of his team’s tropical cyclone work and hopes it means earlier and more accurate warnings; he raises intervening in a wildfire before it escapes control and immediately says that is not something anyone can do today. On energy he is more concrete about where the headroom sits: wind and solar operators already forecast weather to decide what they can sell and at what price, so the gains he expects are further along, in planning grid operations and in predicting demand, including second-order effects like humidity raising the energy cost of heating and cooling air. He also flags the limit on the whole enterprise, which is that improvements come from better data as much as better models, and nobody knows what the ceiling is or which new observation would move it. The honest summary is that this episode establishes a mechanism and a constraint, not a magnitude.