Big Brains Could AI Models Forecast Extreme Weather Events? with Pedram Hassanzadeh
Apr 2, 2026
Pedram Hassanzadeh, Associate Professor of Geophysical Sciences at UChicago who builds AI tools for weather and climate, discusses AI-driven forecasting of extreme heat, hurricanes and floods. He talks about why extremes are so hard to predict, how neural nets can run huge ensembles fast, possibilities for spotting rare “gray swan” events, and real-world impacts like AI forecasts for the Indian monsoon.
AI Snips
Chapters
Transcript
Episode notes
Small Scales Break Big Models
- Weather and climate models struggle because the atmosphere is a multi-scale system where small scales influence large scales.
- Hassanzadeh highlights sub-grid parametrizations (ad hoc functions) at ~10 km resolution as a core source of uncertainty.
How ForecastNet Began
- Pedram started exploring neural nets for weather in 2017 after a student introduced him to deep learning techniques.
- He co-developed ForecastNet with collaborators and trained it on ~40 years of post-1979 satellite-era data, achieving near-physics accuracy much faster.
Cheap Ensembles Improve Extreme Sampling
- AI enables massive ensembles cheaply, improving uncertainty estimates for rare events.
- Hassanzadeh contrasts expensive operational ensembles (~100 members) with neural nets that can produce thousands of runs for better extreme-event sampling.