The AI Weather Model That Uses 200 Times Less Energy Than the World's Best
The European Centre for Medium-Range Weather Forecasts has spent decades running the world’s most sophisticated weather simulations on some of the planet’s largest supercomputers. Their model, the Integrated Forecasting System, crunches through the laws of physics hour by hour to predict what the atmosphere will do next. Last week, ECMWF announced that a machine learning model using 200 times less energy can now match it. In the weeks-ahead range that matters most for disaster planning, it can beat it.
That model is called AIFS-SUBS, and it is ECMWF’s first AI system designed specifically for sub-seasonal forecasting: the awkward middle ground between next week’s weather and next season’s climate. Meteorologists call this the “predictability desert.”
The Gap Between Weather and Climate
Weather forecasts out to about ten days are surprisingly good in 2026. A five-day forecast today is as accurate as a one-day forecast was in 1980. Seasonal outlooks are also reasonably useful. They lean on slowly changing ocean temperatures and sea ice patterns that give broad signals months ahead. But the stretch from roughly day 8 to day 42 is a strange nowhere land. The atmosphere has mostly forgotten its starting conditions, yet the big boundary systems (oceans, ice, land) have not shifted enough to offer anything more than a vague hint. Farmers deciding when to plant, water managers watching reservoir levels, and emergency planners pre-positioning supplies all live in this uncertainty gap.
AIFS-SUBS, described in a paper posted to arXiv last week by Jakob Schloer and colleagues at ECMWF, is designed to make this gap narrower.
Teaching an AI to Look Weeks Ahead
The team did not start from scratch. They adapted AIFS-CRPS, ECMWF’s existing medium-range AI model, making three key changes. First, they stretched the time step from 6 hours to 24 hours. An autoregressive model that predicts the next weather state from the previous one racks up errors with each step, and cutting the step count from about 240 to 42 for a six-week forecast reduces that accumulation substantially. Second, they added data from the stratosphere (the atmospheric layer above roughly 10 kilometers) plus a measure of thermal radiation escaping to space. The stratosphere matters because sudden disruptions there, called sudden stratospheric warmings, can yank the jet stream around and trigger cold snaps at the surface weeks later. Third, they reserved the five-year window from 2007 to 2011 as a completely independent test set. This is a bigger sacrifice than it sounds. For medium-range models, a single year of held-out data provides plenty of independent forecast cases. Sub-seasonal skill hinges on rare events (a given winter might produce only one major stratospheric warming), so you need multiple years to get a meaningful statistical sample.
The architecture itself is a graph neural network with about 230 million parameters, using an encoder-processor-decoder design. It was pretrained on 45 years of ERA5 reanalysis data, then fine-tuned with three-day autoregressive rollouts. The operational version was additionally fine-tuned on recent ECMWF operational analyses, which helps it handle the distribution shift between reanalysis (a best-guess reconstruction of past weather) and the live data that feeds real forecasts.
What the Numbers Say
The team evaluated AIFS-SUBS against the physics-based IFS across the 2007-2011 test period, and the results are striking, not because the AI blows the physics model away, but because it keeps up while using a tiny fraction of the computing budget.
| Metric | AIFS-SUBS vs IFS | Lead Time |
|---|---|---|
| Probabilistic skill (fCRPSS) | Matches or exceeds IFS | Weeks 2-6 |
| Systematic bias | Significantly reduced | Weeks 1-6 |
| MJO forecast skill (OLR correlation > 0.5) | Extended by 8 days | Beyond week 4 |
| Stratospheric warming (SSW) detection | Comparable event frequency | Days 10-36 |
| Energy per forecast | ~200× less than IFS | 46-day run |
| Inference speed | ~920× faster than IFS | 46-day run |
The Madden-Julian Oscillation is a traveling pulse of tropical thunderstorms that circles the globe every 30 to 60 days. Getting its timing right matters because it modulates everything from Indian monsoon onset to Atlantic hurricane formation. AIFS-SUBS extended skillful MJO forecasts by eight days over the IFS. That is the difference between knowing a wet phase will hit Indonesia at week 4 versus week 5.
For stratospheric warmings, the model generated 112 events over the five-year reforecast period, closely matching the IFS’s 98 events. In one specific case, the January 2009 sudden warming, half of AIFS-SUBS’s ensemble members correctly predicted the vortex disruption, while zero IFS members caught it. The surface impacts that followed (cold and dry across Northern Europe, warm and wet in the south) matched the canonical patterns seen in observations.
The team also submitted a variant of their model (trained on ERA5 only) to the AI Weather Quest, an international competition for sub-seasonal forecasting. Across 29 weeks of real-time forecasts from August 2025 to February 2026, AIFS-SUBS scored slightly ahead of the IFS on a variable-averaged basis at weeks 3 and 4, and significantly ahead of the best existing ML competitor, FuXi-S2S.
Why 200× Less Energy Changes the Game
The energy number deserves attention beyond its clickbait appeal. The IFS requires massive CPU clusters for each forecast. AIFS-SUBS runs on GPUs and is roughly 920 times faster in wall-clock time. This is not just an academic exercise in cost savings. It means an operational weather center running AIFS-SUBS could afford to generate ensembles on the order of 1,000 members instead of the 50-ish typical today.
Ensemble size matters because probabilistic forecasts (“there is a 30% chance of a heatwave in week 4”) become more reliable as you add members. The current IFS ensemble spreads its 51 members by perturbing initial conditions and adding random noise to the physics. A 1,000-member AI ensemble would sample the distribution of possible futures far more thoroughly, giving disaster managers and agricultural planners sharper probability estimates at leads where every percentage point translates to real decisions about deploying resources.
The Catch
AIFS-SUBS forecasts the atmosphere in isolation. At sub-seasonal timescales, a large share of predictability comes from the slow churn of the ocean surface, soil moisture, and sea ice. None of these are simulated by the current model. ECMWF’s team acknowledges this and has already published work on coupling the AI atmosphere to ocean models. They also note that the 24-hour time step, while reducing error accumulation, means the model only sees instantaneous snapshots at a single time of day and cannot resolve the diurnal cycle of heating and cooling.
The five-year verification window (2007-2011) is short by the standards of the sub-seasonal community, which typically wants a couple of decades. And because that window sits between two blocks of training data (1979-2006 and 2012-2024), the model may draw on climate trend information from both sides, making the test less stringent than a true out-of-sample future forecast.
Still, six months of live results from the AI Weather Quest, where the model faced genuinely unseen future weather, corroborate the reforecast findings. For the first time, a machine learning model from a major operational center is matching the physics-based gold standard at the timescales where forecasts have historically been least useful.
If 200× cheaper forecasts let weather agencies run ensembles large enough to pinpoint a heatwave’s probability three weeks out instead of one, that changes the calculus for farmers, energy grid operators, and emergency managers in ways that a better five-day forecast never could. The predictability desert is starting to look a little less barren.
📄 Source: Schloer, J., Tietsche, S., Roberts, C.D., et al. (2026). AIFS-SUBS: Extending Data-Driven Forecasting to Sub-Seasonal Timescales. arXiv:2607.05100. European Centre for Medium-Range Weather Forecasts (ECMWF).