Sign up for our newsletter and get the latest big data news and analysis.

HPC Innovation Excellence Award Showcases Physics-based Scientific Discovery

Adam Rupe

A collaboration that includes researchers from NERSC was recently honored with an HPC Innovation Excellence Award for their work on “Physics-Based Unsupervised Discovery of Coherent Structures in Spatiotemporal Systems.” The award was presented in June by Hyperion Research during the ISC19 meeting in Frankfurt, Germany.

Adam Rupe, a PhD student at the University of California, Davis, who has been doing research at NERSC for the last three years, is applying his expertise in physics – specifically, physical principles associated with organized coherent structures – to enable unsupervised discovery of these structures in spatiotemporal (space-time) systems. Through a collaboration with UC Davis physics professor James Crutchfield, Karthik Kashinath and Prabhat of NERSC’s Data & Analytics Services group, and engineers from Intel, the team created the first distributed HPC implementation of a physics-based, data-driven technique known as local causal states. This collaboration, dubbed Project DisCo (Discovery of Coherent structures), led to the HPC Innovation Excellence Award.

With the growing data deluge in climate research, cosmology, materials science, and other science domains, new data-driven methods are required that discover and mathematically describe complex emergent phenomena, uncover the physical and causal mechanisms underlying these phenomena, and better predict these phenomena and how they evolve over time, Rupe explained.

Local causal states, which uncover a system’s spatiotemporal structure by tracking how information is processed locally through space and time, have the potential to do exactly this, and directly from unlabeled data,” Rupe said. “Computational barriers, however, have kept them from reaching this potential on real-world domain science problems.”

As with other data-driven methods, there has long been a disconnect between theoretical development and practical application. DisCo bridges this gap by developing an optimized, distributed HPC implementation of local causal state reconstruction written entirely in Python using standard libraries. “And because it is written in Python, the code can be easily updated as theoretical development continues, giving domain scientists an easy-to-use, high-level API interface to this tool,” Rupe said.

The applications are wide-ranging,” Kashinath added. “From discovering vortices in turbulent flows to extreme weather events in large climate data sets, DisCo is designed to be a tool that scientists can use to understand the underlying physical processes that drive a system.” In addition, DisCo has predictive capability; the local causal states encode a minimal description of the dynamics of the system.

In the work that won the award, DisCo was applied to CAM5.1 climate simulation data on 1,024 Intel Haswell nodes on NERSC’s Cori supercomputer, processing 89.5 TB of data in 6.6 minutes end-to-end. In addition, they obtained 91% weak scaling and 64% strong scaling efficiency. The result was a state-of-the-art segmentation of coherent spatiotemporal structures in complex nonlinear turbulent flows from both observational and simulated data.

Supervised machine learning and deep learning work well only if you have labeled data and you are confident that the labels accurately represent some ground truth. But for these kinds of scientific data sets we essentially do not have ground truth,” Rupe said. “This is why there has been keen interest in unsupervised physics-informed discovery approaches like DisCo.”

This is a video that accompanies the SC19 submission: DisCo: Physics-based Unsupervised Discovery of Coherent Structures in Spatiotemporal Systems. This is a video of unsupervised structural segmentation of vorticity fields from a high-resolution 2D turbulence simulation. Raw fields on the left and causal states corresponding to the structural segmentation on the right. Note that the structural segmentation achieved here is able to extract phenomena of interest including coherent vortices and shear layers in between them. The largest magnitude (darkest color) vorticity regions in the video are elliptic LCS. Each of these elliptic LCS are generally captured by two concentric local causal states; one state for the high-rotation core and one for the boundary. Hyperbolic LCS form between the rotating elliptic LCS and appear as “shrinking and stretching surfaces”. They can be identified in the local causal state field as, for example, narrow white bands between blue and red states (elliptic LCS).

Hyperion Research’s HPC Innovation Excellence Awards are presented twice a year, at the ISC conference in June and the Supercomputing Conference in November. The program’s main goals are to help other users understand the benefits of adopting HPC and justify HPC investments, to demonstrate the value of HPC to funding bodies, to expand public support for increased HPC investments, and to showcase return on investment and scientific success stories involving HPC.

Source: Berkeley Lab

Sign up for our insideHPC Newsletter

Resource Links: