Extreme Weather

Atmospheric blocks, those persistent highs that stall the jet stream for days at a time, can steer storm systems and lock in hazardous weather over a region, from heat domes and drought to flooding, cold air outbreaks, and severe storms. Understanding how these large-scale patterns set up the extreme weather we actually experience on the ground is a key piece of extended-range prediction, and it is a problem where machine learning can really help.

I³: Forecasting blocking extremes for agriculture. Blocking extremes cost the U.S. agricultural industry billions of dollars, yet no forecast model has been built specifically to target blocking events. Through UW-Madison’s Igniting Interdisciplinary Innovation (I³) program, we are part of a highly interdisciplinary team led by Dr. Stephanie Henderson (AOS), spanning atmospheric science, crop and livestock systems, insurance losses, and the Wisconsin State Climatology Office. Our part is the ML model: a vision transformer that learns the spatial structure of blocking development (jet stream structure, ridge amplification) from gridded fields like geopotential height, winds, humidity, and soil moisture, paired with a transformer encoder that tracks how the atmosphere evolves week to week alongside climate indices (MJO, ENSO, NAO, PNA). The model then predicts where blocks are likely to form 1-4 weeks ahead, with explainability built in to understand why.

CIRCS: Cyclones, blocking, and severe convective storms. Over 70% of U.S. tornadoes and convective hazards occur alongside an extratropical cyclone, yet most risk models treat severe storms and cyclones separately. Through the Center for Interdisciplinary Research on Convective Storms (CIRCS), an NSF IUCRC led by NIU and UW-Madison, we are working with Dr. Andrea Lopez Lang and Dr. Stephanie Henderson to build an all-season catalog linking extratropical cyclones, atmospheric blocking, severe storm reports, and even the “null” events that looked dangerous but never produced reports. Our part is the ML side: building on a CIRCS seed project testing whether the time evolution of a block carries predictive signal for cyclone track movement across the Midwest, we use regularized logistic regression, LSTM/transformer sequence models, and SHAP-based explanations to find the drivers of outbreak clustering and seasonal variability, and to build an analog search tool that matches current or forecast setups to similar events from the past.

Our goal is to turn these large-scale signals into actionable, interpretable forecasts of when and where hazardous weather is most likely. Interested in extreme weather and ML? Reach out!