Neural data compression for physics plasma simulation

Jong Choi, Michael Churchill, Qian Gong, Seung-Hoe Ku, Jaemoon Lee, Anand Rangarajan, Sanjay Ranka, Dave Pugmire, CS Chang, Scott Klasky

Keywords: generative model, fusion science, Variational Auto Encoder, Vector Quantized

TL;DR: A generative model for fusion simulation data compression

Abstract:

We present a VAE-based data compression method, called VAe Physics Optimized Reduction (VAPOR), to compress scientific data while preserving physics constraints. VAPOR is based on Vector Quantized Variational Auto Encoder (VQ-VAE) and extended with physics-informed optimization functions and refinement layers, focusing on compressing and reconstructing scientific data with minimum loss of information under physics constraints. We demonstrate VAPOR by using outputs from XGC, a massively parallel fusion simulation code running on the largest supercomputers. Key features of VAPOR are three-fold; i) find a reduced representation of physics data, ii) reconstruct the data with a minimum loss, iii) preserve physics information (e.g., mass, energy, moment conservation).
We discuss challenges in XGC 5D data reconstruction and present our initial experiences and results on how we construct Deep Neural Network (DNN) of VAPOR to optimize the reconstruction quality of XGC data and integrate XGC’s physics constraints.