MIT method trains climate AI with synthetic data

MIT researchers published a technique in the Proceedings of the National Academy of Sciences that accelerates the training of climate emulators using optimized synthetic data. The method generates artificial emission scenarios with high-frequency variations, enabling AI models to generalize better to unseen situations than those trained on realistic historical trajectories. This reduces emulator production time from months or years to days, addressing a major bottleneck in creating up-to-date climate tools for policy planning.