AI falls short on extreme weather forecasting 

Context : The study concludes that while AI models have transformed routine weather prediction, they currently fall short in forecasting unprecedented extreme events, making physics-based models indispensable for reliable early warning systems.

The Study :

• Researchers from Germany and Switzerland compared several leading AI weather models.
• The models assessed included GraphCast, Pangu-Weather and Fuxi.Their performance was compared with the European Centre for Medium-Range Weather Forecasts (ECMWF) model.
• AI models performed well in predicting routine weather and moderate extremes.
• However, they consistently failed to accurately predict unprecedented extreme events.
• AI models are trained using historical weather data and identify patterns from past observations. The study notes that most AI models rely on weather records from 1979–2017.
• AI models struggle with events that fall outside their training data.
• Researchers described this limitation as the “extrapolation problem.”
• Traditional forecasting models use physical laws and mathematical equations to simulate atmospheric processes. Physics-based models can anticipate scenarios that have not occurred before.
• The study tested models using record-breaking heatwaves, cold events and wind events from 2018 and 2020.
• AI models systematically underestimated both the frequency and intensity of these extremes.
• During severe heatwaves, AI models predicted temperatures lower than those actually observed. The greater the departure from historical records, the larger the forecasting error. The study warned that underestimating extremes could weaken disaster preparedness.
• The authors emphasised that early warning systems for heatwaves, storms and extreme weather require the highest possible accuracy.

Mains : 

1. AI and Weather Forecasting

• AI-based weather models are effective for routine forecasting but less reliable for unprecedented extremes.
• Dependence on historical data limits their predictive capability.

2. Limitations of AI Models

• The extrapolation problem prevents AI systems from accurately forecasting conditions beyond their training data.
• Forecast errors increase as weather events become more extreme.

3. Importance of Physics-Based Models

• Traditional models incorporate atmospheric physics and can simulate novel scenarios.
• Physics-based forecasting remains crucial for predicting record-breaking events.

4. Disaster Risk Reduction

• Underestimating extreme weather can compromise early warning systems and preparedness.
• Improving forecasting accuracy is essential for reducing disaster risks under a changing climate.

Weather Forecasting Models & Concepts :

• ECMWF/HRES (European Centre for Medium-Range Weather Forecasts – High Resolution Forecast System)
  • Physics-based global weather model.
  • Considered the gold standard for weather prediction.
  • Widely used by IMD and national meteorological agencies.
• AI-based Weather Models
  • GraphCast (Google), Pangu-Weather (Huawei), FuXi.
  • Trained on ERA5 reanalysis data.
  • Much faster than traditional supercomputer models.
  • Limitation: Depend heavily on historical data.
• Interpolation vs Extrapolation
  • Interpolation: AI Performs well. (Predicting within the range of known data.)
  • Extrapolation: AI Performs poorly! (Predicting beyond known data range)
  • Important limitation in forecasting rare/extreme climate events.