Communications of the ACM
Machine Learning May Be a Game-Changer for Climate Prediction
A major challenge in current climate prediction models is how to accurately represent clouds and their atmospheric heating and moistening. This challenge is behind the wide spread in climate prediction. Yet accurate predictions of global warming in response to increased greenhouse gas concentrations are essential for policy-makers (e.g. the Paris climate agreement).
In "Could Machine Learning Break the Convection Parameterization Deadlock?" published in Geophysical Research Letters, researchers led by Pierre Gentine, associate professor of earth and environmental engineering at Columbia Engineering, demonstrate that machine learning techniques can be used to tackle this issue and better represent clouds in coarse resolution (~100km) climate models, with the potential to narrow the range of prediction.
"This could be a real game-changer for climate prediction," says Gentine, lead author of the paper, and a member of the Earth Institute and the Data Science Institute. "We have large uncertainties in our prediction of the response of the Earth's climate to rising greenhouse gas concentrations. The primary reason is the representation of clouds and how they respond to a change in those gases. Our study shows that machine-learning techniques help us better represent clouds and thus better predict global and regional climate's response to rising greenhouse gas concentrations."
The researchers used an idealized setup (an aquaplanet, or a planet with continents) as a proof of concept for their novel approach to convective parameterization based on machine learning. They trained a deep neural network to learn from a simulation that explicitly represents clouds. The machine-learning representation of clouds, which they named the Cloud Brain (CBRAIN), could skillfully predict many of the cloud heating, moistening, and radiative features that are essential to climate simulation.
"Our approach may open up a new possibility for a future of model representation in climate models, which are data driven and are built 'top-down,' that is, by learning the salient features of the processes we are trying to represent," Gentine says.
The researchers also note that, because global temperature sensitivity to CO2 is strongly linked to cloud representation, CBRAIN may also improve estimates of future temperature. They have tested this in fully coupled climate models and have demonstrated very promising results, showing that this could be used to predict greenhouse gas response.
In addition to Gentine, the Geophysical Research Letters study is authored by Gael Reinaudi of Columbia University, Michael Pritchard and Galen Yacalis of the University of California, Irvine, and Stephan Rasp of LMU Munich.
Pritchard acknowledges funding from the DOE SciDac and Early Career Programs (DE-SC0012152 and DE-SC00-12548) as well as the NSF (AGS-1734164). Stephan Rasp was funded by the German Research Foundation Transregional Collaborative Research Center SFB/TRR 165 "Waves to Weather." Computational resources for SPCAM3 simulations were provided through the NSF Extreme Science and Engineering Discovery Environment under allocation TG-ATM120034.
No entries found