MIT technique reported 2026 to train climate AI emulators faster with less data

Research published in PNAS in the week of Nov 12, 2026 describes a training method that MIT researchers say lets climate emulators adapt using a fraction of the usual data, though testing with full-model scenarios remained future work.

WeeklyTrail · Rowan

Published by WeeklyTrail: 13.11.2026News: 12.11.20261 min read

MIT researchers reported in the week of Nov 12, 2026 a training technique intended to make climate emulators — fast AI surrogate models that mimic full earth system models — quicker to produce and more adaptable, according to the university. The work appeared in the Proceedings of the National Academy of Sciences.

The team said an emulator trained on a much smaller set of carefully optimized synthetic scenarios predicted outcomes of scenarios it had not seen, while traditional emulators they evaluated failed to adapt, including a test involving local pollutant controls. The reported approach requires running a full earth system model once rather than six to eight times. The researchers said they next want to test the method on scenarios generated by a full earth system model and apply it to other model types and machine-learning tasks.

The work was part of the Bringing Computation to the Climate Challenge project and was supported by Schmidt Sciences through the MIT Grand Challenges.