DYffusion: A Dynamics-informed Diffusion Model for Spatiotemporal Forecasting

Authors: Salva Rühling Cachay, Bo Zhao, Hailey Joren, Rose Yu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct an empirical study comparing performance and computational requirements in dynamics forecasting, including state-of-the-art probabilistic methods such as conditional video diffusion models. We find that the proposed approach achieves strong probabilistic forecasts and improves computational efficiency over standard Gaussian diffusion.
Researcher Affiliation Academia Salva Rühling Cachay Bo Zhao Hailey Joren Rose Yu University of California, San Diego {sruhlingcachay, bozhao, hjoren, roseyu}@ucsd.edu
Pseudocode Yes Algorithm 1 DYffusion, Two-stage Training
Open Source Code Yes Code is available at: https://github.com/Rose-STL-Lab/dyffusion
Open Datasets Yes Sea Surface Temperatures (SST): a new dataset based on NOAA OISSTv2 [30], which comes at a daily time-scale. NOAA OI SST V2 High Resolution Dataset data provided by the NOAA PSL, Boulder, Colorado, USA, from their website at https://psl.noaa.gov.
Dataset Splits Yes We train, validate, and test all models for the years 1982-2019, 2020, and 2021, respectively.
Hardware Specification No The paper mentions 'supercomputers' for general numerical simulations but does not specify any particular hardware (GPU/CPU models, memory) used for their own experiments.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies used in the experiments.
Experiment Setup Yes Table 5: The hyperparameters used for each dataset. For the learning rates, we sweep over each value and report the best set of runs based on their validation CRPS computed on 50 samples. DYffusion k refers to the number of artificial diffusion steps used, see Fig. 3. For architectural details, see D.3.