Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere

Authors: Boris Bonev, Thorsten Kurth, Christian Hundt, Jaideep Pathak, Maximilian Baust, Karthik Kashinath, Anima Anandkumar

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply SFNOs to forecasting atmospheric dynamics, and demonstrate stable autoregressive rollouts for a year of simulated time (1,460 steps), while retaining physically plausible dynamics. The SFNO has important implications for machine learning-based simulation of climate dynamics that could eventually help accelerate our response to climate change.
Researcher Affiliation Collaboration 1NVIDIA Corp., Santa Clara, USA 2Caltech, Pasadena, USA.
Pseudocode No No explicit pseudocode or algorithm blocks are provided in the paper.
Open Source Code Yes torch-harmonics and our implementation of the SFNO are available to the public at https://github.com/ NVIDIA/torch-harmonics
Open Datasets Yes The proposed method is applied to the Earth Reanalysis 5 dataset (ERA5) (Hersbach et al., 2020), one of the best estimates of Earth s historical weather and climate over the period 1950-present
Dataset Splits Yes We use 40 years of ERA5 (1979-2018): 1979-2015 is used for training, 2016 and 2017 are used for validation, hyperparameter tuning, and model selection, and 2018 is held out as out-of-sample test set.
Hardware Specification Yes As each autoregressive step takes around 500ms on a NVIDIA A6000 GPU
Software Dependencies Yes To enable our method, we implement torch-harmonics, a library for differentiable Spherical Harmonics written in Py Torch (Paszke et al., 2019).
Experiment Setup Yes Models are trained with the following hyperparameters: 4 (S)FNO blocks, a down-scaling factor of 3, and embedding dimensions of 256.