
NSF Org: |
AGS Division of Atmospheric and Geospace Sciences |
Recipient: |
|
Initial Amendment Date: | August 7, 2024 |
Latest Amendment Date: | August 7, 2024 |
Award Number: | 2427579 |
Award Instrument: | Standard Grant |
Program Manager: |
Eric DeWeaver
edeweave@nsf.gov (703)292-8527 AGS Division of Atmospheric and Geospace Sciences GEO Directorate for Geosciences |
Start Date: | August 1, 2024 |
End Date: | July 31, 2026 (Estimated) |
Total Intended Award Amount: | $202,000.00 |
Total Awarded Amount to Date: | $202,000.00 |
Funds Obligated to Date: |
|
History of Investigator: |
|
Recipient Sponsored Research Office: |
Washington DC US 20011 |
Sponsor Congressional District: |
|
Primary Place of Performance: |
Chicago IL US 60637-1468 |
Primary Place of
Performance Congressional District: |
|
Unique Entity Identifier (UEI): |
|
Parent UEI: |
|
NSF Program(s): | Postdoctoral Fellowships |
Primary Program Source: |
|
Program Reference Code(s): |
|
Program Element Code(s): |
|
Award Agency Code: | 4900 |
Fund Agency Code: | 4900 |
Assistance Listing Number(s): | 47.050 |
ABSTRACT
Climate models use computational methods to solve equations for the physics and dynamics of atmospheric motion, radiative transfer, cloud formation, and other processes that drive the climate system. These physics-based models are remarkably successful but they are also computationally intensive, tracking perhaps hundreds of millions of variables which have to be updated several times in each hour over many decades of simulated time. Computational intensity is a severe limiting factor for efforts to simulate climate change, particularly when the goal is to understand how the frequency and severity of heat waves, rainstorms, droughts, and other extreme events will change as the world warms. Extreme events are rare by definition, thus large ensembles of computationally intensive future climate simulations are required to generate enough extreme events to draw robust conclusions.
The computational intensity of climate models has prompted interest in data-driven methods as a means of generating the large ensembles needed to look at changes in extreme event statistics. These methods typically use climate model output to train a neural network (NN), and while training can be computationally intensive the trained NN can emulate the climate model at a tiny fraction of its computational cost. NNs have shown remarkable skill in making weather forecasts, strongly motivating their use for the study of extreme event statistics.
But it is not clear that current NNs are up to the task. One stumbling block is the "spectral bias" that causes NNs to preferentially learn low frequency signals in training datasets, a deficiency which causes NNs to become unstable when used to generate multiple weeks of simulated weather. A second issue is that NNs are not generally reliable when used to look at conditions which are not represented in their training data, thus it is not clear how well they can represent extreme events which occur only rarely in the training set.
Work performed here seeks to develop NNs capable of capturing extreme event statistics in future climate projections and developing new diagnostics for evaluating them. The work builds on the Principal Investigator's (PI's) previous work developing a NN capable of capturing the transition to chaos in a simple dynamical system using only training data from the pre-transition period when the solution is periodic rather than chaotic. The project also extends the work of Chattopadhyay and Hassanzadeh (2023), who proposed a solution to the spectral bias problem. The project uses a hierarchy of models including a simple two-layer quasi-geostrophic model, intermediate-complexity multi-layer atmospheric models, and the model simulations archived in the Coupled Model Intercomparison Project (CMIP).
The work is of societal as well as scientific interest given concerns over increases in the occurrence and severity of warm-season weather extremes. Efforts to plan for extreme events have traditionally relied on the historical record to estimate return times of weather extremes, but historical observations have limited use if future climate is outside the envelope of conditions recorded in the past. Large ensembles of extreme events generated by NNs trained on climate simulations could thus provide valuable guidance in planning for such events. In addition, the project provides summer research opportunities for undergraduate students at the University of Chicago, where the PI works with the two mentors of this postdoctoral fellowship award, Pedram Hassanzadeh and Tiffany Shaw.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
Please report errors in award information by writing to: awardsearch@nsf.gov.