Measured acceleration · Climate · Physics · Computer science

DLESyM simulates 1,000 years of current climate in hours

A coupled deep-learning Earth-system model that can run long current-climate simulations with much lower compute than conventional models.

Summary

University of Washington researchers reported DLESyM, a deep-learning Earth-system model published in AGU Advances, that simulates up to 1,000 years of current climate and interannual variability. The university summary reports a 12-hour runtime on a single processor for a task estimated at about 90 days on a state-of-the-art supercomputer.

AI role

Used coupled neural-network atmosphere and ocean components to emulate current-climate variability over 100- and 1,000-year rollouts.

Narrative role

This backfills a climate-science measured-acceleration example where AI changes the cost and accessibility of a simulation workflow, not just forecast accuracy.

Caveat

The model targets current-climate variability and does not replace full forced climate projections; the quantitative runtime comparison was reviewed through institutional coverage of the peer-reviewed paper.