Imagining the invisible extreme: MIT's statistical bet and its mathematical price

Forecasting weather events with no historical precedent is one of the hardest problems facing data-driven models, whose uncertainty tends to explode precisely where past examples are missing. To get past that limit, a study published on 20 August 2026 in Nature Communications, titled “Extreme Event Aware (η-) Learning” (article s41467-026-76811-x), proposes constraining training through the statistics of an observable that signals extremity. The authors, doctoral student Kai Chang and professor Themis Sapsis of the MIT Center for Computational Science and Engineering, built that constraint into the loss function as a 1-Wasserstein term. As Kai Chang explained to MIT News, the aim is to “model extreme and unprecedented events that nobody has ever seen before, that are not in the dataset”.

The method was tested on a precipitation downscaling task using the ERA5-Land dataset: hourly total precipitation over the continental United States from 1999 to 2023, aggregated into daily maps. The spatial part of the model was trained on the first six months of the record alone, a window that, according to MIT, contained almost no examples of extreme rainfall. Evaluation was carried out on 9,044 paired high- and low-resolution maps. The approach exists to help planners put numbers on worst cases with no historical precedent, such as the “hundred-year Katrina” cited by Themis Sapsis. The MIT release frames the problem as a question: if the most extreme rainfall ever recorded in New York is 200 millimetres, what kind of storm would produce 300? That is how the question is posed, not a result the model produced.

Reaching beyond the historical record does, however, cost something in pointwise fidelity. According to the figures in the arXiv preprint (which could not be compared with the final Nature Communications version, kept behind a paywall), adding the statistical constraint lowers pointwise fidelity relative to standard training on mean squared error (MSE) alone. Across the full field, root mean square error (RMSE) rises from 2.878 to 3.112 (+8.13%) and the structural similarity index (SSIM) falls from 0.947 to 0.944. On the tail subsets — quantiles ≥0.95, ≥0.975 and ≥0.99 — the RMSE increase runs from 9.21% to 10.30%, while the SSIM loss stays under 0.48%. On the bulk subsets RMSE grows by 6.20-6.92%: accuracy gets worse everywhere, but it gets worse most exactly where the method is supposed to earn its keep. The limitations the authors themselves acknowledge include the impossibility of pinning down the spatial location of extreme events from a scalar observable, a critical dependence on correctly specifying the reference distribution, and the absence of any identification of the physical mechanisms that generate the events.

The work, funded by a Vannevar Bush Faculty Fellowship and the U.S. Air Force Office of Scientific Research, opens the door to theoretical applications in robotic navigation and financial markets as well. The method is demonstrated on one variable, precipitation, and one area, the continental United States, using reanalysis data rather than direct observations; MIT itself writes that extending it to other hazards requires statistics and maps specific to that hazard. There is no independent validation outside the authors' group, and no operational use by civil protection bodies or weather agencies.

The trade-off is stated openly: a few percentage points of accuracy given up everywhere, in exchange for scenarios the historical data does not contain. Whether it is worth it depends on how closely the reference distribution chosen upfront resembles reality — the very point the authors flag as critical.

— Olya