MIT News breed · Wetenschap

MIT Engineers Develop Tool to Simulate Extreme Events Without Past Data

MIT engineers have created a machine-learning algorithm that can generate and map plausible extreme weather events and worst-case scenarios without relying on historical data of past extreme events. This new tool, dubbed Extreme Event Aware (η-learning), learns from available datasets like daily weather records to exclude implausible scenarios and project future extreme events with specific frequencies, intensities, and durations.

Communities often need to understand how extreme events like severe storms, heat waves, or wildfires might unfold to assess risks to infrastructure such as seawalls, power grids, and firefighting resources. However, anticipating these rare outliers is challenging, as traditional risk assessment methods depend on past extreme events to model future worst-case scenarios.

The MIT team's algorithm takes a statistical approach, learning from existing data to generate plausible extreme events that are likely to occur with a given frequency, such as once every 100 years. This allows for the projection of unprecedented events that are riskier than anything previously recorded but still statistically plausible, providing planners with crucial information for infrastructure assessment and reinforcement.

Unlike conventional methods that require training on historical extreme events, the η-learning algorithm combines and learns statistics from point statistics and spatial maps. For instance, by analyzing precipitation maps, it can learn how low-resolution patterns correspond to high-resolution details, enabling it to generate plausible spatial patterns for extreme rainfall events exceeding historical records.

This method can be applied to various fields beyond weather, including robotic navigation and financial markets, to explore complex interactions leading to extreme outcomes like market crashes. The ability to assign probabilities to unprecedented events is becoming critical for national and economic resilience, especially as optimized global systems have little slack to absorb the impact of single extreme events propagating through supply chains and markets.

AI-samenvatting op basis van de bron.