Optimal scenario design for climate emulation
Researchers have developed a novel method to optimize training datasets for machine learning climate models, enhancing their ability to generalize to new scenarios. By using a differentiable Simple Climate Model (SCM), they iteratively adjust training data to maximize emulator skill. This approach, which trains on a single optimized scenario, outperforms emulators trained on multiple standard ScenarioMIP pathways, achieving higher predictive accuracy with less data. The optimized training data allows the emulator to better isolate the distinct physical behaviors of different climate forcing agents like greenhouse gases and aerosols. AI
IMPACT Optimized training data generation could lead to more efficient and accurate climate emulators, reducing the computational cost of climate modeling.