MIT Cuts Climate Emulator Training to 1 ESM Run, Boosting Accuracy on New Scenarios
Updated
Updated · MIT News · Nov 12
MIT Cuts Climate Emulator Training to 1 ESM Run, Boosting Accuracy on New Scenarios
2 articles · Updated · MIT News · Nov 12
Summary
MIT researchers developed a training method that lets climate emulators learn from one full earth system model run instead of the usual six to eight, sharply reducing data needs and production time.
The approach generates optimized synthetic emissions scenarios—rather than realistic ones—and iteratively tunes them until emulator error falls below a target threshold.
Seconds-long emulators trained this way accurately handled unseen cases, including local pollutant-control scenarios, while traditional emulators in the study failed to adapt.
Months-long earth system model runs make current emulators costly and often outdated; the team says faster retraining could improve policy planning for power grids, public health and agriculture.
The work, published this week in the Proceedings of the National Academy of Sciences, will next be tested with full ESM-generated scenarios and extended to other climate and machine-learning models.