Building Data-Driven Surrogate Models at Appropriate Granularity Using High-Resolution Climate Simulation and Coupling Library
DescriptionWhen building a climate model emulator using machine learning, we are concerned about the generalization performance of the surrogate model and the extrapolation of the simulation results. It is very problematic to evaluate an emulator that implicitly includes too many physical processes and has been trained to reproduce only the training data that can exist now. Nor will scientists be able to use the results of such emulators to expand scientific knowledge. We extended the coupling library that combines components of climate models, such as atmospheric and oceanic models, and developed a system for constructing scheme-level surrogate models. With the functionality provided by this coupling library named h3-Open-UTIL/MP, we extract the inputs and outputs of a particular physical scheme under simulation and transform them to different spatiotemporal resolutions as needed. That data is fed on-the-fly to a Python-based machine learning suite. By having the surrogate model behave like a faithful "accelerator" to the original physical model, we can perform long-term simulations with reduced anxiety about unexpected behavior. Also, by constructing a parameterization model of the conventional (low-resolution) climate model based on km-scale simulation, improvements in reproducibility and reduction of computational workload are expected.
TimeWednesday, June 2814:00 - 14:30 CEST
Climate, Weather and Earth Sciences