For decades, weather forecasting has been the domain of massive supercomputers running complex physics-based simulations known as Numerical Weather Prediction (NWP) models. However, a new challenger has emerged: AI-driven models like Google’s GraphCast and Huawei’s Pangu-Weather.
Are these AI models ready to replace traditional physics? The data suggests a more nuanced reality. While AI is making incredible strides, traditional models remain the essential “guardians” of our most extreme weather events.
1. AI is the New King of Tracking
One of the most impressive feats for AI models is their ability to predict the path of major storms. In the category of Tropical Cyclones (TCs), AI models like GraphCast frequently outperform traditional models in multi-day track and landfall location. Because AI is trained on decades of historical data, it excels at recognizing the large-scale patterns that steer these massive storms.
Additionally, AI provides a much more stable “baseline” for everyday weather. For Marginal or Non-Extreme Events, AI models avoid the chaotic “noise” or over-forecasting biases that can sometimes plague physical models, making them highly reliable for your average Tuesday forecast.
2. NWP Still Rules the Extremes
When the weather turns record-breaking, physics still wins. AI models have a tendency toward “smoothing”—they predict the most likely outcome based on history, which often means they under-predict unprecedented extremes.
- Heatwaves & Freezes: For events like Global Heatwaves and Major Freeze Events, traditional NWP models (like the ECMWF IFS) significantly outperform AI. AI models struggle with “domain-extrapolation,” meaning they find it hard to predict temperatures that have rarely or never been seen in the historical training data.
- Storm Intensity: While AI can tell you where a hurricane is going, it often fails to predict how strong it will be. Traditional models are far superior at capturing Tropical Cyclone Intensity and rapid intensification because they calculate the actual physical energy of the storm.
3. The Resolution Gap
Local, high-impact weather still requires the fine-grained detail that only high-resolution regional NWP models can provide.
- Convective Outbreaks: AI models currently lack the resolution to predict fine-scale severe weather features like tornadoes or lightning potential.
- Atmospheric Rivers: While AI captures the large-scale “moisture conveyor belts,” physical models remain the choice for defining localized precipitation impacts along complex mountain ranges.
Summary of Performance
The following chart highlights which model class currently leads across different weather phenomena:
AI vs. Traditional Weather Models: Who Wins Where?
| Metric Category / Phenomenon | Primary Impact Evaluated | Top Performing Model Class |
| Global Heatwaves | Peak temperature magnitude, onset timing, duration | NWP (ECMWF IFS / HRES) |
| Major Freeze Events | Local minimum temperatures, geographic extent | NWP (ECMWF IFS) |
| Tropical Cyclones (TCs) | Track prediction (landfall location & timing) | AI (GraphCast / Pangu) |
| Tropical Cyclone Intensity | Peak wind speed, central pressure, rapid intensification (RI) | NWP (ECMWF IFS / HRES) |
| Atmospheric Rivers | Integrated Vapor Transport (IVT), precise landfall geometry | Tied / Variable |
| Convective Outbreaks | Severe local storms, tornado environments, lightning potential | NWP (High-Resolution Regional) |
| Marginal / Non-Extreme Events (The Forecaster’s Dilemma) | False alarm ratio, skill on average/baseline days | AI Models |
The Verdict: A Hybrid Future
We aren’t looking at a replacement, but a partnership. The future of meteorology likely lies in Hybrid Modeling: using AI for its incredible speed and tracking accuracy, while relying on traditional physical models to anchor our forecasts in the laws of physics during life-threatening extremes.
Whether it’s a record-breaking heatwave or a shifting hurricane track, having both “brains” on the job makes us all a little safer. At Benchmark Labs we combine both physical and AI-based forecasts so our users can benefit from the best of both worlds. Interested in learning more about our solutions? Contact us at info@benchmarklabs.com.
- This post was inspired by the journal article “Extreme Weather Bench: A framework and benchmark for evaluation of high-impact weather” by Amy McGovern et al. (2026) – https://arxiv.org/pdf/2605.01126