ABSTRACT
Machine-learning-based computational methods for structural analysis have been proposed to study colloidal systems. However, most of these methods are based on supervised learning, which suffers from the fundamental difficulty that neural networks cannot correctly discriminate a system that has not been learned in advance. To solve this problem, an anomaly detection method that uses an autoencoder (AE) to distinguish systems with unknown structures was developed. The performance of an AE and a convolutional AE was evaluated, and the properties exhibited by the trained and untrained images in the latent space of the AE with dimensionality reduction were clarified.
Acknowledgments
The author acknowledges the Supercomputer Center at the University of Tokyo for the use of their facilities. This work was supported by JSPS KAKENHI (Grant Number 20K11848).
Disclosure statement
No potential conflict of interest was reported by the author(s).