Intelligent Sampling of Extreme-Scale Turbulence Datasets for Accurate and Efficient Spatiotemporal Model Training
Abstract
With the end of Moore's law and Dennard scaling, efficient training increasingly requires rethinking data volume. Can we train better models with significantly less data via intelligent subsampling? To explore this, we develop SICKLE, a sparse intelligent curation framework for efficient learning, featuring a novel maximum entropy (MaxEnt) sampling approach, scalable training, and energy benchmarking. We compare MaxEnt with random and phase-space sampling on large direct numerical simulation (DNS) datasets of turbulence. Evaluating SICKLE at scale on Frontier, we show that subsampling as a preprocessing step can, in many cases, improve model accuracy and substantially lower energy consumption, with observed reductions of up to 38x.
Keywords
Cite
@article{arxiv.2508.03872,
title = {Intelligent Sampling of Extreme-Scale Turbulence Datasets for Accurate and Efficient Spatiotemporal Model Training},
author = {Wesley Brewer and Murali Meena Gopalakrishnan and Matthias Maiterth and Aditya Kashi and Jong Youl Choi and Pei Zhang and Stephen Nichols and Riccardo Balin and Miles Couchman and Stephen de Bruyn Kops and P. K. Yeung and Daniel Dotson and Rohini Uma-Vaideswaran and Sarp Oral and Feiyi Wang},
journal= {arXiv preprint arXiv:2508.03872},
year = {2025}
}
Comments
13 pages, 9 figures, 2 tables