Introduction
Weather forecasting is critical for preparing for extreme events like heatwaves, cold snaps, and storms, which cause billions in damages annually. For decades, physics-based numerical weather prediction models have been the backbone of forecasting. Recently, AI models have emerged as faster and often more accurate alternatives for routine weather. However, a new study in Science Advances reveals that AI models significantly underperform when predicting record-breaking extreme events. This how-to guide walks you through the key steps to understand why traditional models still outperform AI for these high-stakes forecasts, drawing on the study's findings and broader meteorological principles.
What You Need
- Basic knowledge of weather models: Understand the difference between physics-based (using equations of atmospheric and oceanic processes) and AI-based (using pattern recognition from historical data) approaches.
- Access to study insights: Familiarity with the Science Advances study that tested both model types on thousands of record-breaking hot, cold, and windy events from 2018 and 2020.
- An open mind: Willingness to question the hype around AI and appreciate the strengths and limitations of both methods.
Step 1: Recognize the Strengths of Physics-Based Models
Physics-based models, also called numerical weather prediction (NWP) models, rely on fundamental laws of physics—conservation of mass, energy, and momentum—encoded as differential equations. These equations simulate interactions in the atmosphere and ocean. Because they are rooted in physical reality, they can extrapolate beyond historical data. For example, they can accurately predict a heatwave that has never occurred before, as long as the initial conditions and physics equations are correct. This gives them a critical edge for extreme events where historical analogs may be scarce.
Step 2: Understand How AI Models Learn and Predict
AI weather models, such as those from Google DeepMind or Huawei, use machine learning on decades of historical weather data. They learn patterns: e.g., a certain pressure pattern today often leads to rain tomorrow. During forecasting, the model applies learned correlations without invoking physics. This makes them computationally cheap and often highly accurate for typical weather. However, as study author Prof Sebastian Engelke notes, AI models “depend strongly on the training data” and are “relatively constrained to the range of this dataset.” For record-breaking extremes that fall outside that range, the model can only guess—often underestimating both frequency and intensity.
Step 3: Examine the Study’s Methodology
The study tested both model types on thousands of record-breaking hot, cold, and windy events recorded in 2018 and 2020. They simulated how each model would have forecast these events if run in advance. The comparison focused on two metrics: frequency (how often the model predicted a record-breaking event) and intensity (how severe the predicted event was). By using real historical extremes, the researchers could directly measure each model’s ability to handle rare and novel conditions.
Step 4: Analyze the Key Findings
The results were clear: AI models consistently underestimated both the frequency and intensity of record-breaking events. For example, during a historic heatwave, the AI might predict only a moderate temperature rise, while the physics-based model correctly forecast the extreme. The AI’s reliance on past data made it blind to unprecedented combinations of atmospheric factors. In contrast, physics-based models, armed with physical laws, could simulate novel scenarios even if they had no direct precedent. This finding serves as a “warning shot,” according to the study authors, against replacing traditional models too quickly.
Step 5: Consider the Implications for Forecasting
While AI models excel in many routine forecasts, their failure on extremes has serious consequences. Early warning systems depend on accurate predictions of rare events to save lives and reduce damage. Over-reliance on AI could lead to missed warnings for unprecedented heatwaves or storms. The study suggests a hybrid approach: use physics-based models for extremes and AI for everyday forecasts, leveraging the strengths of each. Governments and agencies should maintain investment in physics-based modeling while cautiously integrating AI as a complementary tool.
Tips
- Don’t count out physics yet: Traditional models are irreplaceable for extreme events—keep them as the backbone of official forecasts.
- Use AI for what it’s good at: Apply AI for short-term, routine predictions where historical patterns are stable.
- Combine approaches: Ensemble forecasting that blends AI and physics-based outputs may offer the best of both worlds.
- Stay updated: As AI training data expands and techniques improve, its performance on extremes may improve—monitor ongoing research.
- Communicate uncertainty: Inform the public that AI models may underestimate extremes, especially for unprecedented events.
By following these steps, you can better appreciate why physics-based models remain essential for extreme weather forecasting, even as AI advances. The key takeaway: never let efficiency replace reliability when lives are at stake.