MIT technique improves climate models with fewer training scenarios

MIT researchers demonstrated a training technique that helped AI climate models predict unfamiliar emissions scenarios using less training data. Their algorithm generates synthetic emissions patterns with rapid variations, allowing a model to extract more information than it gets from conventional, realistic trajectories. In tests, a model trained this way captured local pollution policies that reduce one pollutant while other emissions rise; traditional models failed that test. The researchers say the approach would require one run of a full Earth system model to generate training data, rather than six or eight. They still plan to test the technique using an emissions scenario generated by a full Earth system model.