MIT Trains Climate AI With Less Simulation Data

Researchers at the Massachusetts Institute of Technology (MIT) developed a method that trains climate emulators with less simulation data, according to MIT News. The approach uses a smaller set of optimized synthetic climate scenarios, a fraction of the data usually needed. Emulators are typically trained on six to eight realistic emissions scenarios, each run through a full earth system model.

Climate emulators are AI models that reproduce the behavior of detailed earth system models in seconds. The researchers found that an emulator trained on a smaller set of synthetic scenarios accurately predicted unfamiliar conditions, while conventional emulators trained on realistic emissions trajectories did not adapt as well.

The algorithm creates an initial climate scenario, measures the emulator's error, then repeatedly changes emissions inputs until the error reaches a specified threshold. The resulting scenarios include rapid variations in emissions, which help the emulator learn more about the modeled climate system. The research appears in the Proceedings of the National Academy of Sciences.