MIT develops AI system capable of predicting extreme weather beyond historical records

Researchers at the Massachusetts Institute of Technology (MIT) have developed a new artificial intelligence system that could help governments, engineers and planners prepare for extreme weather events that have never been recorded before.

The new AI approach, developed by MIT engineers Kai Chang and Professor Themis Sapsis, is designed to generate realistic scenarios for rare and unprecedented weather events without needing previous examples of similar disasters in its training data. The technology could help identify risks from storms, floods, heatwaves and wildfires that existing forecasting systems may struggle to predict.

The system, known as Extreme Event Aware learning, or η-learning, works by combining statistical information about how often extreme conditions occur with geographical weather patterns. Instead of simply learning from past disasters, the AI analyses the underlying relationships within available climate data to create possible future scenarios.

Traditional risk models often rely on historical records of major events, using previous storms or floods to estimate what could happen in the future. However, this approach can be limited because the most damaging disasters are rare and there may not be enough historical examples available.

MIT’s researchers say their AI system can produce maps showing what a possible once-in-a-century event could look like, including estimates of its intensity, duration and geographical impact. For example, the system could model a rainfall event more severe than anything previously recorded in a city, helping planners understand how infrastructure might cope.

During testing, the researchers used rainfall data from across the United States. The AI was able to create realistic high-resolution scenarios for extreme precipitation events, including situations where the rainfall levels exceeded those seen in the original dataset.

The technology could have applications in protecting critical infrastructure, including flood defences, electricity networks, transport systems and emergency response planning. By showing how unlikely but possible disasters might unfold, organisations could improve preparations before an event occurs.

MIT researchers say the method could eventually be expanded beyond weather forecasting, with potential uses in areas such as wildfire modelling, financial risk analysis and other fields where rare events can have major consequences.

As climate change increases concerns about more frequent and intense weather events, AI systems that can model scenarios beyond historical experience could become an important tool for improving resilience and protecting communities.