Image across Domains, Experiments, Algorithms and Learning

DOE research across a myriad of science domains are increasingly reliant on image-based data from experiments, but domain scientists continue to struggle to uncover relevant, but hidden, information from digital images. This proposal aims to deliver a new modus operandi for analyzing results of experiments conducted at LBNL and other DOE facilities, providing insight to guide and optimize experiments, in collaboration with colleagues at BES and ASCR. To better exploit the scientific value of a broad array of high resolution, multidimensional datasets, this multi-disciplinary work is designed around a coordinated research effort connecting (1) state-of-the-art data analysis methods with basis on pattern recognition and machine learning; (2) emerging algorithms for dealing with massive data sets; and (3) advances in evolving computer architectures to process the torrent of data. The result will be a set of data science models and new software infrastructure that provides tools that work both “on the factory floor,” as well as workhorse techniques for processing experimental data with supercomputers. In collaboration with colleagues working in energy and health improvements, we have developed computer vision software targeted to scale scientific procedures by reducing the time between experiments, opening more opportunities for more users, and aggregating value to the images obtained at national imaging facilities.