Optimal scenario design for climate emulation

A new method for designing climate model training data shows that a single optimized scenario can outperform six standard emissions pathways in building machine-learning emulators, according to research posted to arXiv this month. The study, published on the preprint server arXiv, introduces a technique that optimizes training datasets to improve the generalization of machine-learning surrogate climate models, also known as emulators [1]. Researchers found that the low structural diversity in existing scenarios commonly used to generate training data places a ceiling on predictive skill [1]. The team used a differentiable Simple Climate Model (SCM) to calculate the sensitivity of emulator loss to perturbations in the training data, iteratively updating the training data to maximize emulator skill [1]. For an SCM, training on one scenario optimized in this fashion outperformed an emulator trained on six standard ScenarioMIP pathways [1]. The emulator achieved higher predictive skill despite training on a smaller dataset, and it successfully isolated distinct physical behaviors of different climate forcing agents — such as greenhouse gases versus aerosols — without requiring single-forcing runs [1]. ScenarioMIP, part of the Coupled Model Intercomparison Project, provides a standardized set of future emissions pathways used across climate research. The new work suggests that the marginal value of generating a small number of dynamically rich scenarios exceeds that of expanding the traditional suite of emissions pathways, particularly in the compute-constrained environment of running full-scale climate models [1]. The researchers further demonstrated that scenarios optimized using an SCM, when used to drive an intermediate-complexity climate model, produce a training dataset that yields a more skillful emulator than training on ScenarioMIP outputs [1]. The method represents a shift in focus from designing model architectures that embed physical constraints — the dominant approach in deep learning for physical systems — to optimizing the data those models learn from [1]. The paper was submitted to arXiv on June 17, 2026, under the physics.ao-ph category [1]. The work appears as the machine-learning community continues to explore ways to improve generalization in surrogate models for physical systems, where training data diversity has emerged as a limiting factor [1].