ClimSim-Online：用于混合机器学习-物理气候模拟的大型多尺度数据集与框架
机器学习
2024-07-10 v6 大气与海洋物理
摘要
现代气候预估因计算限制而缺乏足够的空间和时间分辨率，导致对雷暴等发生在亚分辨率尺度上的关键过程的表征不准确。将物理与机器学习（ML）相结合的混合方法通过将计算密集的高分辨率模拟外包给ML模拟器，提供了更快、更高保真度的气候模拟。然而，这些混合ML-物理模拟需要许多ML专家难以获取的领域特定数据和工作流。作为ClimSim数据集（Yu等人，2024）的扩展，我们提出ClimSim-Online，其还包含用于开发混合ML-物理模拟器的端到端工作流。ClimSim数据集包含57亿对多变量输入/输出向量，捕捉高分辨率、高保真物理对宿主气候模拟器宏观状态的影响。该数据集是全球性的，并以高采样频率跨越十年。我们提供了一个跨平台、容器化的流水线，将ML模型集成到业务气候模拟器中进行混合测试。我们还实现了多种ML基线以及混合基线模拟器，以凸显构建稳定、有技巧的模拟器所面临的ML挑战。数据（https://huggingface.co/datasets/LEAP/ClimSim_high-res）和代码（https://leap-stc.github.io/ClimSim 与 https://github.com/leap-stc/climsim-online）已公开发布，以支持混合ML-物理和高保真气候模拟的发展。
引用
@article{arxiv.2306.08754,
title = {ClimSim-Online: A Large Multi-scale Dataset and Framework for Hybrid ML-physics Climate Emulation},
author = {Sungduk Yu and Zeyuan Hu and Akshay Subramaniam and Walter Hannah and Liran Peng and Jerry Lin and Mohamed Aziz Bhouri and Ritwik Gupta and Björn Lütjens and Justus C. Will and Gunnar Behrens and Julius J. M. Busecke and Nora Loose and Charles I. Stern and Tom Beucler and Bryce Harrop and Helge Heuer and Benjamin R. Hillman and Andrea Jenney and Nana Liu and Alistair White and Tian Zheng and Zhiming Kuang and Fiaz Ahmed and Elizabeth Barnes and Noah D. Brenowitz and Christopher Bretherton and Veronika Eyring and Savannah Ferretti and Nicholas Lutsko and Pierre Gentine and Stephan Mandt and J. David Neelin and Rose Yu and Laure Zanna and Nathan Urban and Janni Yuval and Ryan Abernathey and Pierre Baldi and Wayne Chuang and Yu Huang and Fernando Iglesias-Suarez and Sanket Jantre and Po-Lun Ma and Sara Shamekh and Guang Zhang and Michael Pritchard},
journal= {arXiv preprint arXiv:2306.08754},
year = {2024}
}
备注
This manuscript is an expanded version of our paper that received the Outstanding Paper Award at the NeurIPS 2023 conference