Sampling Hybrid Climate Simulation at Scale to Reliably Improve Machine Learning Parameterization
Download here
Publications
Stress-testing the coupled behavior of hybrid physics-machine learning climate simulations on an unseen, warmer climate
Published in NeurIPS 2023 Climate Change AI Workshop, 2023
This paper samples the coupled behavior of neural network convective parameterizations at scale on an unseen, warmer climate to see if design decisions conducive to better online performance in-distribution do the same out-of-distribution.
Download here
ClimSim: A large multi-scale dataset for hybrid physics-ML climate emulation
Published in NeurIPS 2023 Datasets and Benchmarks Track, 2023
This paper establishes a comprehensive set of datasets and benchmarks for machine learned subgrid parameterizations of convection and radiation in E3SM-MMF. It received the Outstanding Paper award for the Datasets and Benchmarks track at NeurIPS 2023. You can find the official repo for the code here: https://github.com/leap-stc/ClimSim/
Download here
Climate-Invariant Machine Learning
Published in Science Advances, 2023
This paper explores feature transformations that can make machine learned convective parameterizations invariant to distribution-shift in climates.
Download here