[Eoas-seminar] Seminar today by Aaron Hill in EOAS 1044

[eoas-seminar at lists.fsu.edu]

Colleagues,

Aaron Hill, our 3rd MET faculty candidate is speaking today at 3:15 in 
EOAS 1044.  Hopefully you can be there in person. If not the zoom link 
for his talk is

https://fsu.zoom.us/j/95943330514?pwd=VmZkOUN6WXlLR2kwRnBaLy9nbEdDQT09

Meeting ID: 959 4333 0514
Passcode: 632310

Title: Advancing High-Impact Weather Hazard Forecasting with Machine 
Learning

Speaker: Dr. Aaron Hill (Colorado State University)

Abstract: Weather hazards associated with deep convection (e.g., 
excessive rainfall, tornadoes, large hail, and severe wind) are often 
the most costly natural disasters annually. Prediction of these hazards 
is hindered by their localized nature and the inability of sophisticated 
numerical weather prediction (NWP) models to explicitly represent 
processes that result in hazardous weather. Artificial Intelligence (AI) 
and Machine Learning (ML) techniques have emerged recently as 
alternative methods to forecast high-impact weather hazards, and they 
have proven particularly skillful and valuable in their ability to 
explicitly forecast hazard occurrence and location (e.g., 40% chance of 
hail within 40 km of a point). One specific area that AI and ML have 
been used in the meteorology domain is postprocessing of NWP model 
output, taking advantage of the mathematical and statistical properties 
of the ML methods and their ability to process large and complex 
datasets, to create prediction systems capable of generating real-time 
forecasts of weather hazards. One example is the Colorado State 
University Machine Learning Probabilities (CSU-MLP) prediction system 
which is trained to relate historical records of high-impact weather 
with simulated environments from an NWP model. The ML system identifies 
the well-studied synoptic patterns that support high-impact weather and 
can be examined to better understand the ingredients for these events. 
The CSU-MLP has undergone significant development over recent years and 
is now being routinely used in operational forecasting environments, 
including the Weather Prediction Center, Storm Prediction Center, and 
local National Weather Service offices.


Underlying the development of the CSU-MLP, and ML prediction systems 
more generally, is a decision about how to define weather events. 
Whereas tornadoes, severe hail, and severe wind have clear definitions 
(e.g., severe hail is >1” in diameter) and historical records reflect 
these definitions, excessive rainfall and flash flooding are ill 
defined. Does 2 inches of rain in 3 hours produce the same impacts in 
Idaho as it does in Florida? Reports of flash flooding are also 
inconsistently reported across the country due to varying definitions of 
events. As a result, development of the CSU-MLP forecast system has 
considered a number of definitions of excessive rainfall, including 
radar-derived average recurrence intervals (ARIs) and local storm 
reports, to produce a skillful forecast system. This talk will focus on 
these developments and provide a brief overview of the CSU-MLP system 
and forecast skill derived from high-resolution NWP model output as well 
as a coarser global model (i.e., the Global Ensemble Forecast System). 
Interpretability and explainability methods will be introduced to 
demonstrate what can be learned about the forecast problem, including 
how the ML models are learning relevant synoptic and mesoscale dynamics 
that we know exist, which is vital to building trustworthy products. 
Finally, the highlights/challenges of transitioning AI research to 
operational forecast centers will be discussed.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.fsu.edu/pipermail/eoas-seminar/attachments/20230302/cc476dc5/attachment.html>