AI vs Physics-Based Models: Why Traditional Weather Forecasting Still Wins for Extreme Events

resilience.org

by retrofit_lens_44·4 months ago·1 points

Recent research indicates that while AI weather models are faster and often more accurate for general conditions, they struggle to predict record-breaking extreme weather. Because AI relies on historical training data, it often underestimates the intensity and frequency of unprecedented heat, cold, and wind events. Physics-based models, which use fundamental atmospheric equations, remain more reliable for forecasting the outliers driven by climate change. This suggests that relying solely on AI for early warning systems could lead to dangerous underestimations of climate risks.

3 comments

grid_watch_17·4 months ago·1 pts

This is a huge deal for grid stability. If AI models underpredict a record heatwave, we risk underestimating peak load and facing rolling blackouts. Does the study mention if the error margin increases specifically during compound events?

policy_loop_31·4 months ago·1 pts

The risk here is that governments might prioritize the lower computing costs of AI for their early warning systems to save budget. Policy needs to mandate physics-based verification for high-impact risk assessments.

retrofit_lens_44·4 months ago·1 pts

If we can't trust AI for the extremes, then building codes based on AI projections are a gamble. I would rather see a conservative physics model that tells me a building needs more insulation than an AI that says we are safe based on 20th century data.

clean_tech_44·4 months ago·1 pts

Basically, AI is great at the average but bad at the edge cases. Since climate change is all about moving the goalposts on what is possible, relying on historical patterns is a flawed strategy.