Can AI models reliably forecast extreme weather events?

2026-03-16 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Environmental Science & Earth Systems, Mathematics & Computational Sciences, Research Methodology & Innovation · Depth: Intermediate, quick

Summary

Artificial intelligence (AI) models are emerging as a faster alternative to traditional physics-based numerical weather prediction systems, capable of producing a 14-day global forecast two hours earlier. Unlike conventional models that solve complex physical equations, AI models map current conditions to future states using algorithms trained on historical weather data, shifting heavy computing to the training phase. However, concerns exist regarding their reliability in forecasting rare, extreme weather events, as AI systems trained on past data may falter with unprecedented events. While some agencies, like the European Centre for Medium-Range Weather Forecasts, are integrating AI, scientists emphasize the urgent need for rigorous, standardized testing protocols to objectively evaluate AI models' performance on out-of-sample extreme events before widespread adoption by public forecasting agencies.

Key takeaway

For meteorologists and climate scientists evaluating AI integration into operational forecasting, you should prioritize the development and adoption of standardized benchmarking protocols. Your focus must be on rigorously testing AI models against a consensus-driven set of "iconic" extreme weather events that were not part of their training data. This ensures that the speed advantage of AI does not compromise accuracy and reliability when forecasting critical, rare phenomena.

Key insights

AI weather models offer speed but require rigorous, standardized testing for extreme event reliability.

Principles

• Physics-based models maintain validity with climate change.
• AI models may falter with unprecedented events.

Method

The AI Retraining Without Iconic Events (AIRWIE) protocol proposes training AI systems while withholding a designated set of "iconic" extreme events solely for testing, ensuring evaluation against out-of-sample extremes.

In practice

• Benchmark AI models against physics-based counterparts.
• Define minimum standards for hazardous event prediction.

Topics

• AI Weather Forecasting
• Extreme Weather Events
• Numerical Weather Prediction
• Model Evaluation
• Benchmarking Standards

Best for: AI Scientist, Research Scientist, AI Researcher, MLOps Engineer, Policy Maker

Related on AIssential

• Can AI help coastal cities prepare for rising seas and extreme events? — A fast AI emulator accurately predicts extreme storm surges by learning from physics-based models, enhancing coastal flood risk assessment.
• Climate Physicists Face the Ghosts in Their Machines: Clouds — Cloud modeling is a critical bottleneck in climate prediction, driving new AI-physics hybrid approaches.
• Performance Evaluation of GraphCast for Medium-Range Weather Forecasting over Brazil — GraphCast's skill in tropical regions is regime-dependent, excelling in extended range and large-scale moisture.
• How Earth-2 Delivers Global AI Weather Forecasts on Demand — AI models can simulate atmospheric physics directly from observational data, outperforming traditional methods.
• Rigorous uncertainty quantification of probabilistic AI weather forecasts with conformal prediction — Online conformal prediction rigorously calibrates AI weather forecasts, guaranteeing statistical coverage without performance loss.
• FutureSim: Replaying World Events to Evaluate Adaptive Agents — FutureSim evaluates AI agents' adaptive forecasting by replaying real-world events chronologically.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.