MIT researchers unveiled a training method that could speed creation of AI “climate emulators” while improving their reliability, potentially sharpening tools used for policy and planning. The technique, published in the Proceedings of the National Academy of Sciences, generates optimized synthetic emissions scenarios that teach fast surrogate models to mimic complex earth system models with far less data. Instead of relying on a handful of realistic scenarios that can leave emulators brittle, the approach crafts high-variation training inputs that help models generalize to unfamiliar futures, including regional pollutant controls. The result: emulators that run in seconds and better handle out-of-sample conditions, while cutting the need to run expensive full-scale models from multiple times to as little as once. The work, led by Christoher Womack (now at ETH Zurich) with senior author Noelle E. Selin at MIT, aims to provide more trustworthy forecasts for decisions spanning power-grid resilience, public health, and agriculture. Funded through Schmidt Sciences’ MIT Grand Challenges, the team plans to test the method with full earth system outputs and extend it to other climate and machine-learning domains.
Related articles:
Fourier Neural Operator for Parametric Partial Differential Equations
The Scenario Model Intercomparison Project (ScenarioMIP) for CMIP6
