Climate emulators can train on synthetic scenarios: MIT cuts full earth-system model runs needed for training data to one
Synopsis
An MIT team has developed a training-data optimization method that lets a climate emulator trained only on carefully selected synthetic scenarios predict outcomes of scenarios it never saw, while the traditional emulators evaluated failed to adapt to new situations. The method reduces the number of full earth-system model runs needed to generate training data from six to eight down to one, shortening production time for these fast surrogate models. Emulators trained this way also captured a local pollutant-control case in which one pollutant falls while other emissions keep rising.
Interpretation
A climate emulator's training data no longer has to come from realistic emissions pathways; an algorithm can generate scenarios designed specifically to make the emulator more accurate and adaptable. Emulators have typically been trained on six to eight future scenarios based on realistic emissions trajectories, each requiring a run of a full-scale earth system model, making the datasets costly and slow to produce. An emulator trained with a much smaller amount of carefully selected synthetic data accurately predicted outcomes of scenarios it did not see during training, while the traditional emulators evaluated failed to adapt to new situations.
The training pipeline starts from an arbitrary scenario generated by a fast, simple climate model and iteratively adjusts emissions trajectories until the emulator's error in reproducing the full-scale model's results falls to a set threshold. The optimized synthetic scenarios contain high-frequency wiggles and oscillations rather than the smooth trajectories of realistic scenarios, which lets the emulator extract more information about the overall system. In a local pollutant-control scenario where one pollutant falls while other emissions continue rising, the emulator trained with the technique captured the effect, while the traditional emulators evaluated failed this test.
Perspective
The demonstrated scope is training-data generation and prediction testing for a climate emulator, including prediction of scenarios unseen in training and a local pollutant-control case where one pollutant falls while other emissions rise; the researchers next plan to test the technique with an emissions scenario generated by a full earth system model and to apply it to other climate model types and to broader machine learning and computer vision applications.
The synthetic training scenarios are deliberately unlike realistic futures, and the researchers state they do not care much whether training scenarios are realistic; the method has not yet been tested on an emissions scenario generated by a full earth system model, and extension to other climate models and broader machine learning applications remains planned.