One full run trains a faster climate emulator

A full earth system model can take months on a supercomputing cluster to simulate one future, while conventional climate emulators are often trained on six to eight smooth, realistic emissions paths. MIT researchers instead run the full model once and let an algorithm generate optimized scenarios containing rapid increases, decreases and high-frequency oscillations. Those unrealistic inputs force the emulator to learn more of the system response. In tests, the new emulator predicted unseen regional pollution-control combinations that conventional methods missed and could evaluate thousands of scenarios in seconds. The result still needs reproduction across more full earth system models, and the fast surrogate should not be mistaken for the physics model itself.