Multiple machine learning approach to characterize two-dimensional nanoelectronic devices via featurization of charge fluctuation

Abstract

Two-dimensional (2D) layered materials such as graphene, molybdenum disulfide (MoS2), tungsten disulfide (WSe2), and black phosphorus (BP) provide unique opportunities to identify the origin of current fluctuation, mainly arising from their large surface areas compared with those of their bulk counterparts. Among numerous material characterization techniques, nondestructive low-frequency (LF) noise measurement has received significant attention as an ideal tool to identify a dominant scattering origin such as imperfect crystallinity, phonon vibration, interlayer resistance, the Schottky barrier inhomogeneity, and traps and/or defects inside the materials and dielectrics. Despite the benefits of LF noise analysis, however, the large amount of time-resolved current data and the subsequent data fitting process required generally cause difficulty in interpreting LF noise data, thereby limiting its availability and feasibility, particularly for 2D layered van der Waals hetero-structures. Here, we present several model algorithms, which enables the classification of important device information such as the type of channel materials, gate dielectrics, contact metals, and the presence of chemical and electron beam doping using more than 100 LF noise data sets under 32 conditions. Furthermore, we provide insights about the device performance by quantifying the interface trap density and Coulomb scattering parameters. Consequently, the pre-processed 2D array of Mel-frequency cepstral coefficients, converted from the LF noise data of devices undergoing the test, leads to superior efficiency and accuracy compared with that of previous approaches.