Addressing the Data Challenge in at Synchrotrons (#1154)
A. Hexemer1, R. Pandolfi1, D. Kumar1, D. Ushizima1, D. Parkinson1, S. Venkatakrishnan1, C. Tull1, J. Sethian1
1 LBNL, Berkeley, California, United States of America
The advent of high brightness sources, fast detectors and the increasing need of time-resolved experiments in small angle scattering has created an unprecedented data deluge and the needs for combining X-ray science with computer science. Over the last few years we have worked closely with our computational research and supercomputer division to enable the use of supercomputers at scattering beamlines. The dream of such a superfacility would be immediate feedback for scientist during experiments. Such Real-time feedback to scientists during beamtimes is a capability critically needed, however, this dream has not been realized yet. Scattering methods like SAXS and GISAXS (Grazing Incidence Small Angel X-Ray Scattering) generates reciprocal space data that cannot be directly analysed for the underlying material structure. Rather, reverse Monte Carlo and other fitting methods are employed to reverse engineer the sample material. HipGISAXS (High Performance GISAXS) has been developed to run scattering simulations on massively parallel platforms such as the Oak Ridge Supercomputer Titan (OLCF), scalable to thousands of GPUs. Further, with inverse modelling algorithms available in HipGISAXS, such as particle swarm optimization, it can handle a large number of simulations simultaneously during the structure fitting process. In September of 2014 HipGISAXS was used in a real time demonstration that married the SAXS/WAXS beamline at the ALS with the data handling and processing capabilities at NERSC, and simulation capabilities of running at-scale simulations on Titan at OLCF. To accomplish the goal of real time data analysis, we fed the data management and workflow SPOT Suite infrastructure running at NERSC directly with data taken at the beamline. The data was reduced automatically and pushed into CADES at ORN using the high-performance data transfer capabilities of Globus Online. The demo involved the printing of organic photovoltaics using a slot‐die printer installed at the SAXS/WAXS breeamline. Over the span of 3 days many different organic photovoltaics were printed at the beamline and the crystal structure evolution during drying was recorded using GIWAXS. The entire progress of data collection, movement and fitting was monitored on a web dashboard.
Keywords: Synchrotrons, Analysis, Supercomputer
Optimal learning for microstructure reconstruction from high-energy X-ray diffraction data (#1705)
R. Pokharel1, T. Lookman1, P. V. Balachandran1, A. M. Gopakumar1
1 Los Alamos National Laboratory, Los Alamos, New Mexico, United States of America
Increasing data collection rates at 3rd and 4th generation light sources are creating an urgent need for the development of efficient, fast algorithms for reconstruction of microstructural data. Moreover, the use of current data analysis tools is still in their infancy with significant time and resources devoted to manual processing with brute force and inefficient, home-brew scripts. We develop data analysis tools based on learning paradigms for synchrotron data. A key challenge relates to making adequate model predictions to explore the vast landscape defined by the search space. Preliminary results demonstrate that we can reduce the time, effort, and computational resources required for the reconstruction of X-ray diffraction data substantially, while eliminating the manual instrument calibration step. We use Gaussian processes, neural networks, and other machine learning techniques to determine the suitable tools for handling 2D diffraction data for 3D microstructure reconstruction. Our goal is to provide the basis for a software capability for the community that has the potential of changing how experiments and data analyses are currently performed.
Keywords: high-energy X-rays, 3D microstructure, adaptive learning algorithms
Reducing the Combinatorics of Particle Tracking at the High-Luminosity LHC with Machine Learning Techniques and Novel Measurement Techniques (#1162)
L. A. Gray1
1 Fermilab, Scientific Computing Division, Batavia, Illinois, United States of America
The High-Luminosity LHC upgrade poses an immense challenge in terms particle tracking, thousands of tracks corresponding to hundreds of collisions need to be accurately reconstructed with a final rate to disk of around ten kilohertz. This sustained rate is difficult to reach using combinatorial kalman filters, even attempting to harness parallel architectures, due to the underlying combinatorial growth of the problem when multiple viable states are found. To mitigate this behavior, new techniques are required to either better analyze the output of current detectors or to improve the information content of those detectors so that spurious combinations of data are trivially removed. To address the former, machine learning offers a variety of new tools to attack the tracking problem by embedding combinatorics and pattern recognition within neural networks, as well as learning the indicators of fake patterns to produce high-purity output. To address the latter, new advances in silicon detectors allowing precision timing in addition to precision position information are becoming available which can help to significantly sparsify the data being analyzed and reduce confusion and false-positives in new and classical algorithms. The applications of neural networks to particle tracking, particularly addressing reduction in computational complexity and false positive rates. The further application of precision timing to tracking and its impact on reconstruction quality will be addressed as well, including the possibilities coming from the merger of these two techniques.
Keywords: machine learning, precision timing, particle tracking, HL-LHC