← Back to incidents

AI Weather Models Failed to Predict Hurricane Otis Rapid Intensification

Critical

AI weather models including Google's GraphCast failed to predict Hurricane Otis's unprecedented rapid intensification to Category 5, resulting in 46+ deaths and catastrophic damage to Acapulco in October 2023.

Category
Safety Failure
Industry
Government
Status
Resolved
Date Occurred
Oct 25, 2023
Date Reported
Oct 26, 2023
Jurisdiction
International
AI Provider
Google
Model
GraphCast, Pangu-Weather
Application Type
other
Harm Type
physical
Estimated Cost
$12,000,000,000
People Affected
500,000
Human Review in Place
Yes
Litigation Filed
No
weather_predictionrapid_intensificationhurricaneGraphCastPangu-Weatherextreme_eventstraining_data_limitationscatastrophic_failure

Full Description

On October 25, 2023, Hurricane Otis made landfall near Acapulco, Mexico as a Category 5 hurricane with winds of 165 mph, becoming one of the most destructive storms in Mexican Pacific coast history. The hurricane's rapid intensification from a tropical storm to a major hurricane in less than 24 hours caught forecasters completely off guard, with both traditional numerical weather prediction models and advanced AI weather models failing to predict the storm's explosive strengthening. Google DeepMind's GraphCast model, which had been lauded for its superior hurricane track predictions compared to traditional models, failed to forecast Otis's rapid intensification. Similarly, Huawei's Pangu-Weather model and other AI-powered forecasting systems showed the storm remaining at tropical storm intensity. The National Hurricane Center's official forecast on October 24 predicted Otis would reach only Category 1 strength before landfall, a catastrophic underestimate that left residents and authorities unprepared for the unprecedented devastation. The storm's rapid intensification occurred over unusually warm Pacific waters near the Mexican coast, with sea surface temperatures exceeding 30°C (86°F). Otis strengthened from 65 mph winds to 165 mph winds in approximately 24 hours, meeting the definition of 'explosive intensification' with wind speed increases of 100 mph in one day. This rate of intensification was virtually unprecedented in the Eastern Pacific basin's observational record, creating a scenario that fell well outside the training data distributions of AI weather models. The human toll was devastating, with at least 46 confirmed deaths and dozens more missing. Acapulco, a city of nearly one million residents and major tourist destination, suffered catastrophic damage with widespread power outages, collapsed buildings, and complete destruction of the airport and port facilities. Economic losses were estimated at over $12 billion, making Otis one of Mexico's costliest natural disasters. The lack of adequate warning time prevented effective evacuations and emergency preparations, directly contributing to the high casualty count and extensive property damage.

Root Cause

AI weather prediction models were trained on historical data that lacked sufficient examples of rapid intensification events, making them unable to predict extreme and rare meteorological phenomena that fell outside their training distribution.

Mitigation Analysis

Enhanced training datasets with synthetic extreme weather scenarios, ensemble modeling combining AI and physics-based approaches, and improved uncertainty quantification could reduce forecast failures. Real-time model performance monitoring and human expert oversight for anomalous predictions would provide critical safety nets for rare events.

Lessons Learned

The incident highlighted critical limitations of AI weather models when confronted with rare extreme events outside their training distributions. It demonstrated the continued importance of human meteorological expertise and the need for hybrid forecasting approaches that combine AI capabilities with physics-based models and expert judgment.

Sources

Hurricane Otis Discussion Archive
National Hurricane Center · Oct 25, 2023 · regulatory action