MIT Engineers Build AI to Map 100-Year Disasters Without Past Extreme Data

3 articles · Updated · MIT News · Aug 24

MIT engineers unveiled a machine-learning method that generates thousands of statistically plausible worst-case storms, heat waves and wildfires even when historical records contain few or no comparable extremes.
The system learns from ordinary datasets such as daily weather records and spatial maps, then uses point statistics to rule out implausible scenarios while projecting rare events’ likely size, intensity and duration.
In a U.S. precipitation test, researchers used 25 years of hourly rainfall maps but trained the spatial model on just the first 6 months, then produced possible once-in-a-century storms with peaks around 300 millimeters.
The tool, detailed Aug. 20 in Nature Communications, is aimed at planners, insurers and policymakers assessing seawalls, power grids and firefighting capacity, and could also be adapted to robotics and financial-crash modeling.