Technical Area: Data Analytics
We are focused on the development of all aspects of exascale data analytic methods. Efforts include the functional representation of data, using sparse polynomial approximation, low dimensional manifolds, and high order regularizers to enable faster storage, retrieval, and analysis of large datasets. Targeted methods are being developed in sparse storage and retrieval of large data, uncertainty estimates for sparse data representation, fast estimation of data statistics, and importance ranking in streaming data. Some of the techniques considered are described below.
Hierarchical representations of scientific data:
Understanding information in data is a fundamental problem in data analytics; it is magnified with data volume, variety, velocity, and veracity, which most data valued by the DOE possess. Restructuring data in a hierarchical representation intrinsically provides structure to sort information, and enables data discovery and management while reducing both computation and memory loads.
Robust control algorithms for adaptive ML:
We have developed novel application of stochastic PDE theory and tools to problems in scientific data. Efficient stochastic gradient descent algorithm can be controled under the stochastic maximum principle framework to provide robust accellerated training of deep neural networks.
Computational design optimization:
To fuse measured and observed data with computational simulations, we deploy statistical methods to assemble joint posterior distributions that include various data sources modeled according to their resolution/fidelity. This methodology applies to multiscale\multiphysics problems and inverse problems constrained by models with stochastic terms.
Interactive data analytics:
Compressed representation of state variables or quantities of interest, transfered to a common repository, is used to explored in near real-time scientific simulations. We enable users to modify parameters from interactive visualizations to computationally steer their simulation.