Artificial intelligence is helping catastrophe scientists to predict one of the most destructive and least understood severe weather events, the derecho, according to Karen Clark, founder of Karen Clark and Co, the Boston-based catastrophe modelling firm.

A derecho, she explained, “is an organised line of severe thunderstorms that generates damaging winds near or exceeding hurricane force over a wide swath at least 400 miles in length”.

There was a $10 billion derecho that impacted Iowa in 2020, she said.

Until now, derechos could only be defined after the fact during damage surveys. “Scientists have not been able to predict them in advance,” Ms Clark told The Royal Gazette. “KCC scientists are implementing ML [machine learning] techniques to capture derechos in real time as they are occurring.”

KCC’s systems ingest “over 30 gigabytes of data each day … from all the satellites, radar and weather models”. That archive now includes terabytes of high-resolution atmospheric data, which Ms Clark’s team uses to “better predict derechos and tornado outbreaks in real time” as well as “complex local wind patterns for wildfire spread".

While AI can be used for perils with abundant data, Ms Clark noted that it is not applicable for building full hurricane or earthquake models owing to the rarity of those events. “AI and machine-learning techniques require enormous amounts of data … hurricanes and earthquakes are relatively infrequent, [so] there is very little historical data.”

AI’s role in predicting climate change also remains limited, she said, because of the need for a different approach to model training. “AI would need to recognise trends and not just patterns,” she explained. “The future climate is different from past climate, so traditional AI model training techniques are not sufficient.”

She added that the reliability of KCC’s data is maintained through constant updates and rigorous quality checks.

“KCC scientists keep the data current by continuously ingesting daily data,” she said. The company creates detailed “intensity footprints” for hail, wind and tornado events, which insurers rely on to estimate losses in real time.

Of course, AI models are prone to bias, hallucination and error, so Ms Clark said KCC maintains their reliability by comparing model predictions to actual losses months later and “continually scrutinis[ing] those comparisons to make sure there is no bias”.