Highly-scalable, physics-informed GANs for learning solutions of stochastic PDEs

• DOI: 10.48550/arxiv.1910.13444
• Publication Date: 2019-01-01
• Authors: Liu Yang; Sean Treichler; Thorsten Kurth; Keno Fischer; David A. Barajas‐...; Josh Romero; Valentin Churavy; Alexandre M. Tart...; Michael Houston; Prabhat; George Em Karniad...
• Keywords: Symmetric multiprocessor system

Uncertainty quantification for forward and inverse problems is a central challenge across physical biomedical disciplines. We address this the problem of modeling subsurface flow at Hanford Site by combining stochastic computational models with observational data using physics-informed GAN models. The geographic extent, spatial heterogeneity, multiple correlation length scales require training computationally intensive model to thousands dimensions. develop hierarchical scheme exploiting domain parallelism, map discriminators generators GPUs, employ efficient communication schemes ensure stability convergence. developed highly optimized implementation that 27,500 NVIDIA Volta GPUs 4584 nodes on Summit supercomputer 93.1% scaling efficiency, achieving peak sustained half-precision rates 1228 PF/s 1207 PF/s.