Accelerated Mining of 2D Van der Waals Heterojunctions by Integrating Supervised and Unsupervised Learning
An Chen, Zhilong Wang, Xu Zhang, Letian Chen, Hu Xu, Yanqiang Han, Junfei Cai, Zhen Zhou, Jinjin Li
Abstract
The contact resistance of the metal–semiconductor interface at the source and drain increases as the injection of carriers decreases, degrading the device performance. Reducing the Schottky barrier height (SBH) at the metal–semiconductor interface is an effective way of reducing the contact resistance. Therefore, exploring Ohmic contact (no Schottky barrier) or low contact resistance two-dimensional (2D) semiconductor/metal heterojunctions is critical for developing high-performance electronic devices. We generate a comprehensive dataset from the periodic table consisting of 1092 potential 2D semiconductor/metal heterojunctions with good contact performances and demonstrate that the small interfacial dipole and the elimination of localized surface states are essential for designing advanced 2D metal–semiconductor systems with small SBHs. We use integrated supervised and unsupervised learning, as well as first-principles calculations, to screen 6 potential 2D van der Waals metal–semiconductor heterojunctions (BTe–NbSe2, Al2SO–Zn3C2, iAl2SO–Zn3C2, GaSe–NbS2, GaSe–NbSe2, and GeSe–VS2) with Ohmic contact and high tunneling probabilities from 1092 candidates. The proposed method takes less than 5 s to execute and is far superior to traditional first-principles calculations in both time and cost, demonstrating the superiority of using machine learning for screening materials and that unsupervised assisted algorithm can alleviate the problem of data scarcity to predict the behaviors of complex dynamical systems.