High-speed classification of coherent X-ray diffraction Patterns on the K computer for high-resolution single biomolecule imaging

Tokuhisa, Atsushi*; Arai, Junya*; Jochi, Yasumasa*; Ono, Yoshiyuki*; Kameyama, Toyohisa*; Yamamoto, Keiji*; Hatanaka, Masayuki*; Gerofi, B.*; Shimada, Akio*; Kurokawa, Motoyoshi*; Shoji, Fumiyoshi*; Okada, Kensuke*; Sugimoto, Takashi*; Yamaga, Mitsuhiro*; Tanaka, Ryotaro*; Yokokawa, Mitsuo*; Hori, Atsushi*; Ishikawa, Yutaka*; Hatsui, Takaki*; Go, Nobuhiro*

Single-particle coherent X-ray diffraction imaging using X-ray free electron laser has potential to reveal a three-dimensional structure of a biological supra-molecule at sub-nano meter resolution. In order to realize this method, it is necessary to analyze as many as one million noisy X-ray diffraction patterns, each for an unknown random target orientation. To cope with the severe quantum noise we need to classify patterns according to their similarities and average similar patterns to improve the S/N ratio. We developed a high-speed scalable scheme to carry out classification on the K computer, a 10PFLOPS supercomputer at RIKEN Advanced Institute for Computational Science. It is designed to work on the real time basis with the experimental diffraction pattern collection at the X-ray free electron laser facility SACLA so that the result of classification can be feed-backed to optimize experimental parameters during the experiment. We report the present status of our effort of developing the system and also a result of application to a set of simulated diffraction patterns. We succeeded in classification of about one million diffraction patterns by running 255 separate one-hour jobs on 385-node mode.