Optimal scenario design for climate emulation.

An emulator trained on a single SCM scenario optimized via differentiable loss sensitivity outperforms emulators trained on six standard ScenarioMIP pathways, despite using fewer training samples.

• Why it matters: Limited structural diversity in commonly used climate training scenarios can cap emulator predictive skill, motivating dataset-level optimization rather than only architecture changes.
• What they did: The method iteratively perturbs and updates training data by computing how emulator loss changes under training-data perturbations using a differentiable simple climate model (SCM).
• The result: Optimized scenarios isolate distinct physical behaviors of different forcing agents (e.g., greenhouse gases vs. aerosols) without single-forcing runs, and SCM-optimized scenarios also improve emulator skill when used to train an intermediate-complexity climate model.

AI summary of the abstract · may contain errors · not medical advice · read the paper.