MIT researchers introduced a machine learning framework called Extreme Event Aware Learning, or η-learning, designed to predict rare and extreme events without requiring extreme occurrences in training datasets, according to a peer-reviewed paper published by Nature Communications.
Existing data-driven approaches struggle with rare phenomena because severe events happen infrequently and remain expensive to simulate. While conventional methods produce accurate predictions in quiescent regimes, they face high epistemic uncertainty in extreme conditions unless multiple extremes appear in the training or sampling stages.
The η-learning method reduces uncertainty across uncharted regimes by applying statistical regularization during training. The framework enforces the statistics of an observable indicative of extremeness, derived either from unlabeled data or qualitative knowledge. This constraint allows models to match observed baseline data while remaining consistent with prescribed observable statistics, enabling the generation of unprecedented extreme events.
Optimal-transport-based theoretical results offer justification for the framework and establish its optimality properties. Authors Kai Chang and Themistoklis P. Sapsis evaluated the framework using numerical experiments on prototype systems and real-world precipitation downscaling tasks.
Both authors conducted the research at the Massachusetts Institute of Technology within the Department of Mechanical Engineering and the Center for Computational Science and Engineering. The project received funding from the Air Force Office of Scientific Research under grant FA9550-23-1-0517 and the Vannevar Bush Faculty Fellowship under grant VBFF N000142512059, with theoretical feedback contributed by Jennifer Zheng of Stanford ICME. Nature Communications accepted the paper on August 5, 2026, following its initial submission on February 8, 2026.
