Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision

Abstract A promising approach to improve climate‐model simulations is to replace traditional subgrid parameterizations based on simplified physical models by machine learning algorithms that are data‐driven. However, neural networks (NNs) often lead to instabilities and climate drift when coupled to...

Full description

Saved in:
Bibliographic Details
Main Authors: Janni Yuval, Paul A. O'Gorman, Chris N. Hill
Format: Article
Language:English
Published: Wiley 2021-03-01
Series:Geophysical Research Letters
Subjects:
Online Access:https://doi.org/10.1029/2020GL091363
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850281611661148160
author Janni Yuval
Paul A. O'Gorman
Chris N. Hill
author_facet Janni Yuval
Paul A. O'Gorman
Chris N. Hill
author_sort Janni Yuval
collection DOAJ
description Abstract A promising approach to improve climate‐model simulations is to replace traditional subgrid parameterizations based on simplified physical models by machine learning algorithms that are data‐driven. However, neural networks (NNs) often lead to instabilities and climate drift when coupled to an atmospheric model. Here, we learn an NN parameterization from a high‐resolution atmospheric simulation in an idealized domain by accurately calculating subgrid terms through coarse graining. The NN parameterization has a structure that ensures physical constraints are respected, such as by predicting subgrid fluxes instead of tendencies. The NN parameterization leads to stable simulations that replicate the climate of the high‐resolution simulation with similar accuracy to a successful random‐forest parameterization while needing far less memory. We find that the simulations are stable for different horizontal resolutions and a variety of NN architectures, and that an NN with substantially reduced numerical precision could decrease computational costs without affecting the quality of simulations.
format Article
id doaj-art-a8022d9b88a14cd08c85a7a910447dd2
institution OA Journals
issn 0094-8276
1944-8007
language English
publishDate 2021-03-01
publisher Wiley
record_format Article
series Geophysical Research Letters
spelling doaj-art-a8022d9b88a14cd08c85a7a910447dd22025-08-20T01:48:15ZengWileyGeophysical Research Letters0094-82761944-80072021-03-01486n/an/a10.1029/2020GL091363Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced PrecisionJanni Yuval0Paul A. O'Gorman1Chris N. Hill2Department of Earth, Atmospheric and Planetary Sciences Massachusetts Institute of Technology Cambridge MA USADepartment of Earth, Atmospheric and Planetary Sciences Massachusetts Institute of Technology Cambridge MA USADepartment of Earth, Atmospheric and Planetary Sciences Massachusetts Institute of Technology Cambridge MA USAAbstract A promising approach to improve climate‐model simulations is to replace traditional subgrid parameterizations based on simplified physical models by machine learning algorithms that are data‐driven. However, neural networks (NNs) often lead to instabilities and climate drift when coupled to an atmospheric model. Here, we learn an NN parameterization from a high‐resolution atmospheric simulation in an idealized domain by accurately calculating subgrid terms through coarse graining. The NN parameterization has a structure that ensures physical constraints are respected, such as by predicting subgrid fluxes instead of tendencies. The NN parameterization leads to stable simulations that replicate the climate of the high‐resolution simulation with similar accuracy to a successful random‐forest parameterization while needing far less memory. We find that the simulations are stable for different horizontal resolutions and a variety of NN architectures, and that an NN with substantially reduced numerical precision could decrease computational costs without affecting the quality of simulations.https://doi.org/10.1029/2020GL091363Atmospheric modelingConvectionmachine learningparameterizationsubgrid physics
spellingShingle Janni Yuval
Paul A. O'Gorman
Chris N. Hill
Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision
Geophysical Research Letters
Atmospheric modeling
Convection
machine learning
parameterization
subgrid physics
title Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision
title_full Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision
title_fullStr Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision
title_full_unstemmed Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision
title_short Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision
title_sort use of neural networks for stable accurate and physically consistent parameterization of subgrid atmospheric processes with good performance at reduced precision
topic Atmospheric modeling
Convection
machine learning
parameterization
subgrid physics
url https://doi.org/10.1029/2020GL091363
work_keys_str_mv AT janniyuval useofneuralnetworksforstableaccurateandphysicallyconsistentparameterizationofsubgridatmosphericprocesseswithgoodperformanceatreducedprecision
AT paulaogorman useofneuralnetworksforstableaccurateandphysicallyconsistentparameterizationofsubgridatmosphericprocesseswithgoodperformanceatreducedprecision
AT chrisnhill useofneuralnetworksforstableaccurateandphysicallyconsistentparameterizationofsubgridatmosphericprocesseswithgoodperformanceatreducedprecision