APPLICATION OF SPARSE DECOMPOSITION, DIMENSION REDUCTION, AND EMERGENT FEATURE RECOVERY TECHNIQUES TO LEARNING FROM MASSIVE PHYSICS BASED SIMULATION

ABSTRACT: We investigate in the context of massive-scale CFD simulation datasets, modern information extraction and feature discovery methods recently developed in the context of analyzing complex, high-dimensional sensor data. We include four task threads: (1) Dimensionality-reduction to examine correlation of emergent feature spaces with physical phenomena. (2) Identification and tracking of coherent features sets across time-steps. (3) Machine learning of correlations between emergent structures and measurable scalar parameters like lift and drag. (4) Use of simulation data to discover signature characteristics detectable by practical physical sensors. With significant experience in applying the subject algorithms to other classes of data, we can offer a cost-effective, high-return program by evaluating several interrelated approaches in single, a multi-thread effort. We exploit cutting edge mathematics related to compressive sensing and emerging concepts in data-adaptive, sparse image and signal processing. SR2 Group is teaming with Prof. Wojciech Czaja of The Norbert Wiener Center For Harmonic Analysis and Applications at the University of Maryland, and Dr. Elias Balaras of The George Washington University. We bring significant experience in DOD and civilian applications of these algorithms coupled with world-class expertise on large-scale CFD simulations. BENEFIT: This effort is synergistic with in-house algorithm development programs and will contribute to advancing capability across a spectrum of applications. Specific methods addressed in this proposal have two obvious markets: (1) Government applications, including DOD multi-physics simulations (2) and commercial CFD software users. CFD in itself spans numerous industries, including aerospace, automotive, naval architecture, and even sports and recreational equipment. Related market opportunities exist in providing a combinations of proprietary software and signature processing expertise to help solve difficult problems in mapping of emergent features in CFD or other complex multi-physics, chemical, or life-science models to macroscopic phenomena.