Steven Greybush
Pennsylvania State University, PA·Earth & Environmental Science
Thunderstorms bring heavy rain, damaging winds, large hail and tornadoes. Forecasting these hazards depends on knowing exactly where and when a storm will first form, a moment called convection initiation (CI).
- Fields
Earth & Environmental Science
Computer Science
- Topics
convection initiation
thunderstorms
nowcasting
- Grants
- 1 known grant, through Aug 2027
What they work on
- Design and optimize a deep learning nowcasting system for convection initiation that combines forecasting, diagnosis and generation of storm scenarios, trained on high-resolution satellite, radar and surface weather data.
- Quantify how predictable thunderstorm initiation is and how uncertain the AI forecasts are, using new uncertainty methods that produce well-tuned ensembles, and compare them against traditional ensemble numerical weather prediction.
- Uncover the physical and dynamical mechanisms that control storm initiation by using physically informed explainable AI that links realistic changes in physical processes to changes in the forecasts.
- Explore pairing AI with data assimilation, which blends observations into weather models, by using AI as the ensemble forecast model and to estimate background error covariance, a measure of how model errors are related.
More about the research
Today's storm-resolving computer weather models do well at following a storm once it exists, but they struggle to predict that first spark. Artificial intelligence (AI) models can learn patterns in space and time across many scales. However, it is often unclear what physical patterns they have actually learned and why their forecasts improve.
Why it matters
Better predictions of where thunderstorms will first form could give forecasters more lead time for flash floods, hail, damaging winds and tornadoes, helping communities and emergency managers prepare. The CI benchmark dataset and the uncertainty estimates for AI forecasts could also help weather services judge when machine-learning predictions can be trusted alongside traditional weather models.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Chengsi Liu
University of Oklahoma, OK·Earth & Environmental Science
Develop a machine-learning-guided adaptive inflation method that corrects under-dispersive ensembles (forecast sets that are too similar to one another) in storm regions during radar data assimilation.
Michael French
Stony Brook University, NY·Earth & Environmental Science
Establish a firm link between updraft area at middle and low levels and the occurrence of tornadoes and large hail in more than 100 supercells, using an automated algorithm that detects ZDR columns in dual-polarization radar data.
Kate Musgrave
Colorado State University-Fort Collins, CO·Earth & Environmental Science
Connect guest artificial-intelligence researchers and their hurricane models to NOAA's Hurricane and Ocean Testbed (HOT), with the project team serving as intermediary between AI developers and the operational testbed.
The school
Pennsylvania State University: 60 labs on PI Pages, $1.3B a year on research, 743 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.