Application of ML and KDE for efficient data treatment on single crystal diffraction data

単結晶回折データに対する機械学習およびカーネル密度推定の適用

鬼柳 亮嗣  ; 大原 高志   ; 中尾 朗子*; 宗像 孝司*; 石川 喜久*; 森山 健太郎*

Kiyanagi, Ryoji; Ohara, Takashi; Nakao, Akiko*; Munakata, Koji*; Ishikawa, Yoshihisa*; Moriyama, Kentaro*

"SENJU" is a TOF-Laue neutron single crystal diffractometer installed at J-PARC/MLF in Japan. The experiments conducted at SENJU can be categorized into two types, one is for a standard structure analyses where a large number of Bragg reflections are collected, and the other is for a search for superlattice reflections including magnetic reflections where new reflections, which typically are weak, are searched within the observed 3D reciprocal space. For these types of data, some mathematical methods are applied in order to efficiently treat the data. (1) "Application of machine learning to Bragg reflection selection"; For a standard structure analysis, typically some hundreds, or often more than ten thousand, of Bragg reflections are measured. The issue is that the measured data have to be checked before fed into a structure analysis software. In order to efficiently check the data, machine learning was adapted. The model obtained by the convolutional neural network was able to distinguish "good" and "bad" data with high accuracy. Further improvement is envisaged by increasing number of the training data. (2) "Application of kernel density estimation to reciprocal map"; For a search for superlattice reflections, the expected reflections are typically very weak. Hence longer exposure is needed and, often, even after the long exposure, the reflections could be blurry. In order to enhance the chance to find new reflections, the kernel density estimation (KDE) method was applied to a measured reciprocal map data. The visibility of the reflections is found to be greatly enhanced by the application of KDE by roughly five times, which means that the application of KDE can be equal to five times longer exposure.