Energy-Aware Earth System Science: A Community Roadmap
In the authors' words
Earth system models underpin the evidence that informs net-zero policy, yet their energy demands are rarely considered when judging model quality. Using archived output from 159 models spanning five generations of the Coupled Model Intercomparison Project (CMIP), we show that model error has fallen at a constant rate of 13.5 per cent per doubling of compute since 1995. Halving CMIP6 error would require 28 times more compute. Improvements are uneven, with sea-level pressure and surface air temperature errors declining 3.5-fold and 2.7-fold since CMIP1, but precipitation error by only 7 per cent. Machine-learning weather models match or exceed operational forecast skill at a fraction of the runtime, but none publishes measured energy, and including training changes how they rank. We propose a skill-energy framework that combines an empirical Pareto front with a ranking score. These findings argue for standardised reporting of energy use and performance ahead of CMIP8.
In plain words AI summary
Earth system models guide climate policy, but their energy use is rarely weighed when judging quality. Model error has dropped 13.5 percent each time computing power doubled since 1995. Cutting CMIP6 error in half would need 28 times more computing power. Some errors improved more than others: sea-level pressure and surface temperature fell 3.5-fold and 2.7-fold, but precipitation only 7 percent. The authors propose a skill-energy framework and call for standard energy and performance reporting before CMIP8.
Written by Kimi K2.5 using only the abstract above. Jev checked that every sentence is supported by the abstract sentences it cites, and removed any that were not. Point at a sentence to highlight its source.
Appeared: Monday, October 19. arXiv. Preprint, not yet peer-reviewed.