Blog · Field notes · Oct 21, 2026 · 5 min
Learned weather models are now operational, faster and often more skilful than physics on ordinary days, and weaker on record-breaking ones. The habits that made them trustworthy, daily scoring against what happened and honesty about extremes, are what simulation of lives and economies needs next.
Weather is the first domain where learned simulation went from research to the forecast people actually receive. In 2025 Europe's leading forecasting centre put two machine-learning models into operation, a foundation model of the Earth system appeared in Nature, and generative models began to emulate the climate at kilometre scale. A careful study in 2026 then showed exactly where those models still fall short: the record-breaking day. Both halves of that story carry a lesson for anyone simulating something harder than the atmosphere, such as a population or an economy.
The European Centre for Medium-Range Weather Forecasts began running its deterministic learned model, AIFS Single, operationally on 25 February 2025, and its ensemble, AIFS ENS, on 1 July 2025. Both were upgraded to a second version on 12 May 2026. ECMWF reports that the 51-member AI ensemble outperforms its physics-based ensemble on many measures, with gains of up to 20% for surface temperature, while running more than ten times faster and using about a thousandth of the energy. It runs at 31 km against the physics ensemble's 9 km, and ECMWF keeps the physics system for high-resolution fields and coupled Earth-system processes.
Microsoft's Aurora, published in Nature on 29 May 2025, was trained on more than a million hours of diverse geophysical data and, in the authors' results, outperformed operational forecasts of air quality, ocean waves, tropical cyclone tracks and high-resolution weather at orders of magnitude lower computational cost. The broader claim is that one pre-trained model can be fine-tuned for many Earth-system tasks cheaply.
Google's WeatherNext 2, released on 17 November 2025, generates hundreds of possible outcomes from a single starting point, each in under a minute on one TPU, and surpasses its predecessor on 99.9% of variables and lead times from 0 to 15 days. It now feeds forecasts in Search, Gemini, Pixel Weather and the Maps Platform.
For climate, NVIDIA's cBottle, announced in June 2025 as part of Earth-2, is a generative model that produces global atmospheric states at kilometre scale, conditioned on inputs such as sea surface temperature, time of day and day of year. And Europe's Destination Earth, now in its third phase, runs a climate twin that replays record heatwaves and floods in a world 2°C warmer.
The sharpest test of these models came from Zhang, Fischer, Zscheischler and Engelke. Published in Science Advances in 2026 and summarised in KIT's press release, it compared GraphCast, Pangu-Weather and FuXi with ECMWF's physics-based high-resolution forecast on record-breaking heat, cold and wind. The learned models did well on ordinary weather and worse on records: they underestimated both how often and how strongly records were broken, and "the greater the exceedance of the record of their training data, the larger the underestimation". The authors' advice is direct: "For high-risk applications, one should not rely solely on AI."
That is the general behaviour of learned models, stated precisely. Inside the range of their training data they are fast and often more skilful. Beyond it, they shrink towards what they have seen. The days that matter most are often the days furthest outside the record.
Weather had three things few other domains have. It has a shared record of the atmosphere on which models are trained and compared with each other. It has a habit of scoring each forecast against what actually happened, so a model's skill is measured rather than asserted. And it keeps physics models in the loop, as ECMWF's decision to retain its physics ensemble for high-resolution fields and coupled processes shows, and as the extremes study recommends for high-risk days.
Lives and economies have weaker versions of all three. Their records are fragmented across censuses, surveys and registers. Their predictions are rarely scored, because the period that would test them is years away. And the equivalent of a physics model, the mechanism by which births, partnerships, migration and shocks produce what we later measure, is exactly what a simulation of lives has to supply.
The step from emulating the atmosphere to simulating people is a step from a field governed by known equations to a system whose rules have to be built from measured rates and then tested. The lessons carry across directly. Score forecasts on periods nobody had seen. Report the error honestly, including the misses. Treat shocks as events with their own chances, because the extreme year is where averages fail. And say where a model holds and where it does not, as the extremes study does for weather.
Living Worlds simulates environments, land, water and air, alongside the people who live in them, and its lifetime forecasts are scored as weather is: frozen, then compared once with figures from years nobody had read.
Frozen on data up to 2005, Living Worlds forecast women's life expectancy twenty years ahead within 0.92 years on 2025 figures from four national offices (Met, women only), and men's within 1.09 years across six, with its 80% intervals holding all six countries (Met). For European regions ten years ahead it came within 0.64 years for women and 0.33 for men on a window nobody had read (Met).
The misses are where weather's lesson about extremes shows up. Fitted on 1950-1999 and scored on 2019 across 155 countries, a forecast came within 2.06 years for women and 2.42 for men against a target of 1.0, and most of what remained was shocks: wars, state collapse and HIV (Not met). Frozen in 2004 and scored on 2024 figures for 37 countries, the forecast expected the longest-lived countries to keep gaining as they had, and they did not (Not met). Treating wars, collapses and epidemics as events with their own chances gave women honest 80% intervals, 79% coverage across 155 countries, while men's intervals covered 65% (Met for women, not for men). Every one of these was registered before scoring.
Sources
Blog
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.
Living Worlds' life expectancy forecasts, frozen twenty years back and scored once on years no run had seen: met for women and for men on the latest national figures, not met against 2024 national figures or across 155 countries, with honest intervals for women and a stacked UN model that did not help.
A feeder model that learned from Living Worlds' checked simulation first, then from a handful of real measurements, predicted 30 unmeasured low-voltage feeders better than the same model on real measurements alone; power through 40 unmetered feeders beat physics and a learned model; thermal comfort in fifteen countries came close but was not met.