Scaling the superemulator for ClimateSet : extending machine learning capabilities to 36 climate models

Abstract

Recently, ClimateSet has been introduced, providing ML engineers with consistent climate model data from the Input4MIPs and CMIP6 Archives for tasks like predicting specific climate scenarios, downscaling, or multi-model emulation (Superemulation). While Superemulation can speed up multi-model forecasting and potentially enhance the quality of the climate predictions, scaling up to 36 models was not possible due to performance and scalability problems. This project designs a Climate Model Superemulation that can emulate 36 Climate Models within one machine learning model using a Convolutional LSTM, a U-Net, and a transformer-based architecture, while profiling various performance metrics. The scaling impact of optimization techniques has been analyzed, along with potential sustainability hand- and footprints in ML-based Climate Emulation. Performance experiments are conducted using a 6-model Superemulation, analyzing scaling impacts before and after optimization on a 36-model Superemulation. The results show significant reductions in GPU/CPU utilization, increased accuracy, and model speed-ups of up to 70%, along with decreased environmental footprints. This work enables the Superemulation of all 36 CMIP6 Climate Models in ClimateSet, supporting future work on Superemulation approaches and discussing the usefulness and sustainability of ML-based Emulation approaches.