Total: 1
A central challenge in climate science and applied mathematics is developing data-driven models of multiscale systems that capture not only stationary statistics but also responses to external perturbations. Current neural climate emulators seek to resolve the atmosphere-ocean system in all its complexity but often fail to reproduce forced responses, limiting their use in causal studies such as Green's function experiments. Response theory offers a rigorous framework to explore these limitations, moving beyond stationary statistics, probing causal mechanisms, and guiding model design. Using a simplified multiscale system, we argue that the skill of purely data-driven models in reproducing perturbed statistics depends critically on (i) identifying the proper variables to model, including spatial and temporal scales, and (ii) accurate stochastic parameterizations of unresolved processes. For studies of low-frequency climate dynamics, these insights motivate reduced-order neural models, tailored to specific processes, as valuable alternatives to general-purpose emulators. We illustrate this perspective with a reduced-order neural stochastic model designed to investigate responses of radiative fluxes to surface temperature perturbations. The model largely reproduces stationary statistics and enables the study of forced responses in both ensemble mean and variance, shifting the study of climate feedbacks from single trajectories to responses of probability distributions. These results underscore reduced-order models, augmented with modern neural networks and evaluated within the framework of response theory, as a practical strategy for causal inference and attribution studies in large-scale climate dynamics.