How AI weather models work

Among the routes Saily WR draws for a passage, three come from models unlike the others: ECMWF's AIFS, DWD's AICON and NOAA's AIGFS. For seventy years, a forecast meant solving the physics of the atmosphere; their wind comes instead from neural networks that learnt the weather from its own past. The words "artificial intelligence" invite either blind faith or a shrug, and neither helps a skipper weighing a passage. To know how much weight their routes deserve, it helps to open the box and see what goes on inside.

A classic weather model is physics put to work. It cuts the atmosphere into boxes, about 9 km across and stacked 137 high for ECMWF's IFS, and in each one applies the laws that govern the air: Newton's laws for its motion, the conservation of its mass, its energy and its water. Step after step, a few minutes of weather at a time, it works out how wind, pressure, temperature and moisture change. What is smaller than a box, a cumulus cloud, the turbulence over the waves, cannot be computed and is represented by simplified rules, the parametrisations, where much of a model's error hides. A forecast to fifteen days keeps a large share of a supercomputer busy.

Every forecast, classic or not, starts from a picture of the present, the analysis. ECMWF receives some 800 million observations a day, from satellites, weather balloons, aircraft, ships, drifting buoys and weather stations, and keeps about sixty million once checked, yet they leave wide gaps, above all over the oceans. Data assimilation fills them: it takes a short forecast of the physical model and draws it towards each observation, as far as that observation deserves to be trusted, until it strikes the best balance between what the model expected and what was measured. This step, the most delicate of the whole chain, matters for what follows: none of the AI models in Saily WR performs it. Each one starts from the analysis of a physical model.

The leap of the AI models is to replace the equations with experience. Picture a navigator who had studied every weather map of the last forty years, six hours apart, and learnt by heart how each one turned into the next: shown the map of now and the map of six hours ago, they could sketch the one six hours on. That is exactly the question a neural network is set: given the state of the atmosphere now and one step earlier, predict it one step later, six hours ahead for AIFS and AIGFS, three for AICON. No law of physics is written into it: everything it knows, it found in its examples.

Its textbook is a reanalysis: the weather of the past rebuilt with a single model and a single method of assimilation, so that every year is drawn alike. ERA5, ECMWF's reanalysis, gives the state of the atmosphere every hour since 1940, on a grid of a quarter of a degree. AIFS learnt from its years 1979 to 2022, then from ECMWF's recent analyses; GraphCast, the Google DeepMind network at the heart of AIGFS, from ERA5 and ECMWF's analyses too, before NOAA fine-tuned it on its own; AICON from fifteen years of a reanalysis DWD made with ICON. Training means showing the network millions of pairs of states, measuring each time how far its guess falls from what happened, and adjusting its millions of parameters, the weights of its connections, to fall a little closer. It is done once for each version, about a week on 64 graphics processors for AIFS, and the model is then used as it is.

How does a network take in a whole planet? Not as a flat picture: the Earth is round, and on a grid of latitudes and longitudes the meridians crowd together at the poles. All three gather the weather onto a mesh of points spread evenly over the sphere and linked to their neighbours, a graph. GraphCast builds it by dividing, again and again, the faces of an icosahedron, a die with twenty faces, until it has some 40,000 nodes; AICON draws it from ICON's own triangles. In GraphCast and AICON, the nodes then pass messages to one another: short links carry the local weather, long links let information cross an ocean in a single step. AICON weighs those messages by attention, a mechanism by which each point learns which of its neighbours matter most; AIFS, whose heart is a transformer rather than a graph, applies attention to windows that slide over its mesh. At the end, the result is spread back onto the grid of the forecast.

A network that only knows how to look one step ahead reaches days ahead by feeding on itself: its forecast for six hours becomes the starting point of the next step, and so on, 64 times for the sixteen days of AIGFS. This chaining, called autoregression, is the weak link: an error made early is carried into every step after it, and can grow. To hold it in check, the networks are trained not only on one step but also on several chained together, so that they learn to correct their own drift. That training has a hidden cost, which the next paragraphs come to: DWD chose to train AICON on a single three-hour step precisely to keep the detail that longer training rubs away.

Once trained, these models are strikingly frugal. A ten-day AIFS forecast takes a few minutes on a single graphics processor, where the IFS needs a large share of a supercomputer, and ECMWF puts the energy saved at about a thousand times. That changes what a forecasting centre can afford: running fifty forecasts at once to measure the uncertainty, as AIFS ENS does with its 51 members, costs little. But the saving stops at the forecast itself. The analysis it starts from still comes from the physical system and its observations, at full cost: the AI models make the forecast cheaper, not the knowledge of the present.

Their best-known flaw comes from the way they are marked. During training, the error is most often measured as a mean squared error, the average of the squared gaps between forecast and truth. Five days ahead, the atmosphere has several possible futures: a low may pass north of Ireland, or south. Faced with that doubt, the answer that keeps a squared error smallest is neither track but their average, a shallower low somewhere in between. The network learns to hedge, and its fields grow smoother the further it looks: wind peaks are blunted, fronts softened, storms weakened. ECMWF notes that AIFS makes tropical cyclones less intense than the IFS does, and DWD's forecasters found that AICON underestimates strong winds.

Research is busy with the cure. Ensembles trained on another score, the continuous ranked probability score or CRPS, which rewards a set of forecasts whose spread matches the real uncertainty, keep each member sharp instead of letting them collapse into one smooth average: that is the road of AIFS ENS. The version 1.1 of AIGFS, since July, measures its errors on the patterns of each field, scale by scale, rather than point by point, to blur less far ahead. AICON, trained on a single short step, comes out less blurry, by DWD's own measure, than the deterministic models trained on ERA5. None of these cures is complete, and Saily WR reads a single forecast from each model: in a blow, the wind of an AI model is still best read as probably too light.

Where AIFS and AIGFS shine is the large picture several days out, the part of a forecast that decides an ocean passage: where the lows and highs will lie, when a front will arrive, which way a tropical storm will go. In its original form, started from ECMWF's analyses, GraphCast beat the best deterministic forecast of its time on 90% of the 1,380 targets it was checked against; ECMWF found AIFS ahead of its physics-based IFS on many measures, by up to 20%, tropical cyclone tracks among them. AICON, built for the weather near the surface, does best on the wind of the first days, and gives way to ICON on the large systems beyond about two days. These are scores of positions and patterns, not of the strength of the wind in a squall.

Their limits come from what they learnt. A network knows only the weather of its textbook: a storm stronger than any in its decades of reanalysis lies outside its experience, and nothing in it guarantees respect for the laws of physics beyond what its examples taught it. The first version of AIFS, badly weighted on the moisture of the soil during its training, forecast spots of rain out of nowhere. Their grids are coarse, about 30 km for AIFS and AIGFS, and AICON's 13 km triangles reach Saily WR on a quarter-degree grid like theirs: no sea breeze, no acceleration off a cape, no wind shadow behind an island. And they do not replace the physical models, they lean on them: without the analysis of the IFS, GFS or ICON, they would have no starting point.

That dependence is precisely what makes them useful at sea. Each AI model has a physical twin that starts from the very same analysis: AIFS with ECMWF's IFS, AICON with ICON, AIGFS with GFS. Where the routes of two twins part, they set out from the same weather, so the gap comes from the models alone: a sign that the forecast is uncertain, or that the AI model is smoothing it. Say you are in A Coruña, waiting for a window to cross the Bay of Biscay in four days: if AIFS and the IFS agree, and GFS, which starts from another analysis, agrees with them, the window stands on firm ground; if AIFS alone shows a gentler low, it is most likely smoothing it. Read AIFS and AIGFS for the where and the when of the large systems, AICON for the wind of the first days, the physical models for how hard it blows, and their disagreement as the measure of your doubt.