The climate projections that inform decisions about power grids, public health and farming come from earth system models so detailed that a single scenario can take months on a supercomputer cluster. Scientists get around that with emulators, fast AI stand-ins that mimic the big models and run in seconds. MIT researchers now report a way to train those emulators on much less data, and to make them better at situations they have not seen.
The usual approach, MIT News explains, is to train an emulator on six to eight realistic future emissions pathways, each of which has to be run through a full earth system model first. That is slow and expensive, and the team had earlier found it can produce emulators that fail on unfamiliar scenarios. Their new algorithm instead builds a synthetic training set designed purely to teach the emulator well. It starts from an arbitrary scenario produced by a fast, simple climate model, measures how well the emulator reproduces it, and then adjusts the emissions in the training data, for instance a sharp rise in carbon dioxide in 2050 or a quick drop in methane in 2037, until the error falls below a threshold.
The resulting scenarios look nothing like a plausible future. They are full of rapid swings and oscillations, which the lead author, Christopher Womack, says let the emulator extract much more information about how the system responds. In tests reported in the Proceedings of the National Academy of Sciences, an emulator trained this way predicted scenarios outside its training set, including local pollution controls that cut one pollutant while others kept rising, a case the traditional emulators evaluated failed. A user would need to run the full model once rather than six or eight times.
The idea will be familiar to anyone who works on machine learning: choose training data for what it teaches, not for how realistic it looks. It echoes the broader shift in science towards fast learned models, from Google DeepMind's weather work to the wider family of AI emulators, where the value lies in testing thousands of scenarios cheaply.
The limits are significant. The training scenarios in this study came from a simple climate model, and the researchers say testing with a scenario generated by a full earth system model is still future work. Results are for the emulators and cases evaluated in the paper. Whether policymakers will trust a tool trained on deliberately unrealistic data is a separate question the research does not settle.