MIT researchers describe an algorithm that generates extreme weather scenarios without extreme data
A method that extrapolates from ordinary conditions shows how models can reason about rare events their training data never contained.
Researchers at MIT have described an algorithm that produces plausible scenarios for extreme events that have never appeared in the historical record, MIT News reported on 24 August 2026. The work comes from Kai Chang, a PhD student at the MIT Center for Computational Science and Engineering, and Themis Sapsis, the William I. Koch Professor of Mechanical and Ocean Engineering. Their paper appeared in Nature Communications on 20 August.
The method is called Extreme Event Aware learning, abbreviated as η-learning. It is designed to answer a question that archives cannot: what would an event look like if it were more severe than anything ever measured? Rather than waiting for a rare storm to occur, the approach combines two kinds of information. One is point statistics, meaning probability estimates for quantities such as the maximum rainfall a location can receive. The other is spatial maps that describe how conditions are distributed across a region. By learning the relationship between low-resolution and high-resolution data under ordinary conditions, the model extrapolates to intensities it was never shown.
The researchers evaluated the approach on 25 years of hourly precipitation data covering the continental United States. In one demonstration, they generated maps of plausible rainfall events reaching 300 millimetres over New York City, above the roughly 200-millimetre maximum in the historical record. The outputs characterise the potential duration, intensity and geographic footprint of an event rather than issuing a single forecast.
Sapsis summarised the motivation as asking what a once-in-a-century Katrina would look like and how bad it could be. The team says the same machinery could be applied to floods, wildfires, robotic navigation and financial market crashes. The research was supported by a Vannevar Bush Faculty Fellowship and the U.S. Air Force Office of Scientific Research.
Why it matters
Machine learning models are usually judged on how faithfully they reproduce their training distribution, which is the wrong yardstick when the goal is planning for rare disasters. This work points toward models built to reason about the tails of a distribution rather than its centre. If the approach generalises across domains, infrastructure planners and risk teams could stress-test systems against events the archive has never contained. It also suggests that statistical structure, not just data volume, is becoming a central design lever in scientific AI.
Machine learning models are usually judged on how faithfully they reproduce their training distribution, which is the wrong yardstick when the goal is planning for rare disasters.
Point statistics
Probability estimates for a quantity at a location, such as the chance rainfall exceeds a given level
Spatial maps
Observed fields showing how conditions are distributed across a region under normal conditions
Learn low-to-high resolution relationship
The model learns how coarse summaries relate to fine spatial detail
Extrapolate to unseen intensity
Push the statistics beyond the historical maximum and let the learned structure fill in the map
Stress-test infrastructure
Use the generated duration, intensity and footprint as inputs to planning, not as a forecast
What you can learn from this
- Interpolation versus extrapolation. A standard supervised model learns a mapping from inputs to outputs across the range of examples it has seen, and it tends to be reliable only inside that range. Ask it about a rainfall total larger than any in the training set and it has no examples to anchor the answer, so its output is a guess shaped by whatever function it happened to fit. That is why unprecedented events are so hard for conventional models: the region of input space that matters most is exactly the region with no data.
- Tail statistics and extreme value theory. Point statistics describe how likely it is that a quantity exceeds a given threshold at a single location, and a branch of statistics called extreme value theory shows that the maxima of many processes follow a small family of characteristic shapes. Fitting that shape to observed annual maxima lets you estimate a 100-year value even from a 25-year record, because the tail is constrained by theory rather than by counting events. The MIT work leans on this idea: the probability of a severe intensity can be estimated even when that intensity has never been observed. The weakness is that the fit is only as good as the assumption that the underlying process is stable over time.
- Scenario generators are not forecasts. A forecast tries to say what will happen at a specific time, while a generative model samples plausible states from a learned distribution. The η-learning outputs are maps of what a severe event could look like, so they are judged on coverage and physical plausibility rather than on hit rate. This distinction changes how you use the result: a scenario is an input to a stress test, not a prediction to act on. Confusing the two is a common mistake when generative tools enter risk work.
- Learning across resolutions. Coarse data such as regional totals is plentiful and stable, while fine-grained maps are rarer and noisier. A model that learns how fine detail relates to coarse summaries under normal conditions can be steered by pushing the coarse summary into extreme territory and letting the learned relationship fill in the spatial pattern. This is the same idea behind super-resolution in image processing, applied to physical fields. It works because spatial structure, such as how a storm's rain band is organised, tends to persist even as intensity changes.
- Validating the unvalidatable. Because no ground truth exists for an event that has never happened, validation must be indirect. Researchers typically hold out the most extreme observed cases, train on the rest, and check whether the model would have anticipated them. They also test physical consistency, such as whether generated rainfall respects the spatial correlation patterns seen in real data.
We teach this
How to use this in practice
- Fit an extreme value distribution to real data. Download a daily precipitation or temperature series for a station near you from a public meteorological archive, take the maximum value for each year, and fit a generalised extreme value distribution with
scipy.stats.genextreme.fit. Then compute the 100-year return level from the fitted parameters. Done looks like a histogram of annual maxima with the fitted curve overlaid and a printed return level that exceeds the largest observed value. - Run a held-out-tail experiment. Take any regression dataset you have, train a simple model on rows whose target is below the 90th percentile, and evaluate on the rows above it. Record the error on the held-out tail alongside the error on an ordinary random split. Try a second model type, such as a tree ensemble against a linear model, and note that tree-based models cannot predict beyond the largest target they have seen. Done looks like a two-row table showing how much worse extrapolation is than interpolation for the same model, which is the problem this research is trying to solve.
- Draw the pipeline. On paper or in a diagramming tool, sketch the flow from observed data to point statistics and spatial maps, into a model that learns the low-to-high resolution relationship, and out to a generated extreme scenario. Annotate which boxes are constrained by observed data and which rely on extrapolation. Done looks like a one-page diagram you could use to explain the method to a colleague in two minutes.
- Write a scenario for a system you own. Pick a service, a home lab or a personal project, and note the largest load or input it has ever handled. Describe what a 1.5x event would look like, name the component you expect to fail first, and state the threshold at which it fails. Done looks like a half-page document with three components, three thresholds and one mitigation for each. Revisit it after the next incident to see whether your threshold was right.
Sources
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report