Sandia_Natl_Labs_FY19_LDRD_Annual_SAND2020-3752 R_2_S

FY19 ANNUAL REPORT

Parallel tensor decomposition for massive, heterogeneous, incomplete data. Tensor decom- position is a fundamental tool for unsupervised machine learning and understanding complex data sets. Unlike deep neural nets and related tools, tensor decomposition can identify structure in data in a way that’s transparent to users and can work with very limited data. The LDRD-developed Generalized Canonical Polyadic (GCP) tensor decomposition is especially unique in its ability to handle binary, count, and nonnegative data, in contrast to standard approaches geared to normal-distributed data. Sandia recently developed GenTen, open source C++ software for GCP tensor decomposition, that achieves high performance on multiple platforms including many-core central

processing units and graphical processing units. Data science applications for this software include sensor monitoring, cyber security, treaty verification, and signal processing, where it identifies latent structure within data, enabling anomaly detection, process monitoring, and scientific discovery. 

GCP tensor decomposition enables alternate objective functions and more analysis options including binary and nonnegative data. (Image by Tamara Kolda)

Stochastic optimization to enhance resiliency and response strategies in critical infrastructure. The loss of critical infrastructure services, such as electric power grids, water systems, and communication networks, can be caused by natural hazards or intentional acts.  The U.S. critical infrastructures must be designed so that they are robust under abnormal conditions such as line faults, generator failure, water contamination, and computing network intrusions. Sandia developed software for optimal design and operation of critical power systems that considers uncertainty, discrete decisions, and nonlinear physics associated with the critical

infrastructure.  These techniques have been published in six journal articles, and the capabilities for mixed-integer nonlinear programming and electrical grid optimization have been incorporated into two open source software packages based on Pyomo ( CORMIN and EGRET ). Follow-on work is being funded by ARPA-E and the DOE Office of Fossil Energy (FE).

The CORMIN and EGRET software was applied to two exemplary problems including the design and operation of nonlinear power systems and stochastic sensor placement. (Image by Carl Laird)

18

LABORATORY DIRECTED RESEARCH & DEVELOPMENT

Made with FlippingBook Learn more on our blog