GC13V-07 Accounting for Memory and Explainability in Emulating Dynamical Systems

We incorporate memory into a climate emulator by using linear response functions represented as a combination of spatial and temporal modes. Previous work has shown that data-driven techniques with memory, such as response functions, can improve upon simple pattern scaling approaches in emulating climate variables. Purely data-driven approaches often lack explainability and are computationally intensive, generally requiring a large scenario ensemble where each member has a large number of spatial dimensions. Using a smaller ensemble results in emulators which either lack accuracy or are numerically unstable. Our novel technique enhances our understanding of system dynamics, guarantees stability, and reduces the spatial dimensions required for emulator diagnosis, decreasing computational complexity without sacrificing accuracy or explainability. We first demonstrate how response functions can be derived directly through modal decomposition or indirectly through a least-squares fit of leading-order time scales. Then, we apply this technique to emulate the ensemble mean response of pedagogical systems, providing insights into system dynamics. We conclude by showing its skill in predicting climate variables when trained on full scale Earth System Models. Like other linear response functions, we can accurately represent the climate response within non-monotonic warming scenarios (e.g. SSP119/126). However, while response functions have typically been derived from perturbation experiments (e.g. 4xCO2), our methodology is applicable to more general scenarios (e.g. 1pctCO2) and can be used to characterize the response to spatially dependent forcings, such as aerosols. While our emulator predicts the ensemble mean response of a given system, future work using the same method could integrate stochastic terms to allow for the generation of new scenario ensembles.