Application of the AI2 Climate Emulator to E3SMv2's Global Atmosphere Model, With a Focus on Precipitation Fidelity

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 3
• Main Authors: Duncan, James P. C., Wu, Elynn, Golaz, Jean‐Christophe, Caldwell, Peter M., Watt‐Meyer, Oliver, Clark, Spencer K., McGibbon, Jeremy, Dresdner, Gideon, Kashinath, Karthik, Bonev, Boris, Pritchard, Michael S., Bretherton, Christopher S.
• Format: Journal Article
• Language: English
• Published: 01.09.2024
• Subjects: climate modeling; global atmosphere model; machine learning; precipitation; tropics
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000136

Can the current successes of global machine learning‐based weather simulators be generalized beyond 2‐week forecasts to stable and accurate multiyear runs? The recently developed AI2 Climate Emulator (ACE) suggests this is feasible, based upon 10‐year simulations with a network trained on output from a physics‐based global atmosphere model using a grid spacing of approximately 110 km and forced by a repeating annual cycle of sea‐surface temperature. Here we show that ACE, without modification, can be trained to emulate another major atmospheric model, EAMv2, run at a comparable grid spacing for at least 10 years with similarly small climate biases—a prerequisite to wider applicability. With an analysis that combines multiple temporal, spatial, and frequency domain perspectives, we show that ACE faithfully represents the spatiotemporal structure of EAMv2 precipitation and related variables. Finally, we show that a pretrained ACE network is able to adapt to a new global climate model simulation data set with 10× ${\times} $ fewer training steps than when starting from random initialization, all while still maintaining low levels of climate bias. Further analysis of these fine‐tuning experiments reveal ACE's intriguing ability to interpolate between distinct global climate models. Plain Language Summary Traditional methods to predict the weather use mathematical models of the Earth's atmosphere that are costly to run. However, “data‐driven” weather prediction methods, which learn to predict future weather directly from data on past weather, have come to match or even beat traditional methods and do so with much less running cost. In contrast to weather prediction where the goal is to predict the weather in the near future, in climate modeling the goal is to study the Earth's long‐term weather trends under different possible future scenarios for many years into the future. Until the introduction of the AI2 Climate Emulator (ACE), a recent data‐driven method for climate modeling, no data‐driven method could match traditional climate models. In this work we test ACE's climate modeling skills and find that it is able to faithfully mimic a traditional model of the climate when looking at patterns of rainfall around the globe and in the tropics. After learning to mimic one climate model, ACE is able to adapt to a new climate model in much less time. These results show that ACE has the potential to simulate the Earth's climate under many more scenarios and with much lower cost than ever before. Key Points The AI2 Climate Emulator (ACE) yields an accurate climate when trained on EAMv2, E3SMv2's global atmosphere model ACE captures the space‐time organization of EAMv2 precipitation well, with a much smaller time‐mean bias than EAMv2's observational bias Fine‐tuning experiments show linearity of climate biases when interpolating between ACE‐FV3GFS and ACE‐EAMv2