Machine learning for predicting physical parameters of atom-vacancy defects from low-frequency noise in few-atom layer MoS2

Abstract

The rapid advancement of deep learning (DL) has significantly expanded its application across various fields, including physics and materials science. DL facilitates the prediction of physical parameters of materials without the need for time-consuming and labor-intensive experiments. Atom-vacancy defects, present in various materials responsible for intriguing physical phenomena, exhibit diverse physical parameters and are a prime example of such applications. These tendencies are even more pronounced in two-dimensional (2D) (atomically) thin van der Waals layers. However, the correlation between different types of atomic defects and their properties remains an area of active exploration. Herein, we propose an innovative DL model and predict the structural parameters of atom-vacancy defects solely from low-frequency noises, arising from stochastic capture and emission of charges through them, in few-atom layer semiconductors molybdenum disulfides (MoS2). The predicted parameters show good agreement with experimental results. This methodology enables the accurate extraction of structural parameters of various materials, significantly reducing error rates by suppressing excess learning and demonstrating the high potential of DL.