Accelerated Data-Driven Accurate Positioning of the Band Edges of MXenes Avanish Mishra, †,¶ Swanti Satsangi, †,¶ Arunkumar Chitteth Rajan, † Hiroshi Mizuseki, ‡ Kwang-Ryeol Lee, ‡ and Abhishek K. Singh* ,† † Materials Research Centre, Indian Institute of Science, Bangalore 560012, India ‡ Computational Science Research Center, Korea Institute of Science and Technology (KIST), Seoul 02792, Republic of Korea * S Supporting Information ABSTRACT: Functionalized MXene has emerged a promising class of two-dimensional materials having more than tens of thousands of compounds, whose uses may range from electronics to energy applications. Other than the band gap, these properties rely on the accurate position of the band edges. Hence, to synthesize MXenes for various applications, a prior knowledge of the accurate position of their band edges at an absolute scale is essential; computing these with conventional methods would take years for all the MXenes. Here, we develop a machine learning model for positioning the band edges with GW level of accuracy having a minimum root-mean-squared error of 0.12 eV. An intuitive model is proposed based on the combination of Perdew-Burke-Ernzerhof band edge and vacuum potential having a correlation of 0.93 with GW band edges. These models can be utilized to identify MXenes for a desired application in an accelerated manner. M Xenes (M n+1 X n ; M, group IIIB to VIB; X, {C, N}; and n =1-3) 1-6 are a vast class of two-dimensional (2D) materials exfoliated from corresponding MAX phases, 6-8 which get functionalized because of unsaturated surface charges. Previously, more than 25 000 functionalized MXenes (MM′XTT′; T: F, O, OH, etc.) have been generated, 9 which are metallic or semiconducting depending upon surface termination (T). MXenes possess variability in their properties and are considered promising for electronic, photovoltaic, and photocatalytic applications. 10-12 These properties rely on the absolute position of their band edges. For example, in the case of electronic devices, the direction of charge transfer and the height of the Schottky barrier (band offset) depend on the position of band edges. Also, the water-splitting ability of photocatalytic materials, semiconductor heterojunctions in a laser, and separation and migration of photogenerated charge carriers in photovoltaics depend on the position of conduction and valence band edges (Figure S1). Determining the accurate position of band edges for all the MXenes would be time-consuming as it requires knowledge of ionization potential (IP) and electron affinity (EA), which involve the calculation of vacuum potential (ϕ) and band extremas. 13 These are calculated using a density functional theory (DFT)-based approach, wherein the external potential is determined with an additive constant using ground-state charge density. Hence, the calculated eigenvalues are arbitrary and need to be referenced to a uniform scale. Using ϕ as a reference, the band edges are positioned to the absolute scale (Figure S2). A local or semilocal functional-based approach within DFT always underestimates the band gap. 14 Further- more, it fails to predict the accurate position of the valence band edge. 15,16 Attempts are made to improve the accuracy in the estimation of band edges at an absolute scale using hybrid methods such as Heyd-Scuseria-Ernzerhof (HSE06) 17-19 and Becke, three-parameter, Lee-Yang-Parr (B3LYP). 20 Although these hybrid methods improve the band gap accuracy, the position of the band edges differs from the experimental value. 15,16 This problem can be solved to a large extent by using a many-body perturbation theory-based GW approach. 21 However, the GW method is computationally very expensive. Therefore, it would be extremely time-consuming to estimate the accurate position of the band edges for all the MXenes. Recently, machine learning (ML) has been shown to be promising for faster and accurate prediction of unknown structures 22-24 and properties 9,25-28 of large materials data sets within a reasonable time. In this Letter, we have developed ML models to accurately position the band edges of MXenes at an absolute scale. The ML models are developed by mapping GW valence or conduction band edges (E V GW /E C GW ) to easily accessible properties of the MXenes. The Gaussian process regression (GPR)-based ML model is developed to predict GW band edges with a root-mean-squared error (rmse) of 0.12 eV. Moreover, by analyzing the correlation of features, an intuitive model is proposed, which is able to capture the physical origin behind the shift of reference level and unravel the role of surface functionalization in controlling it. These ML models Received: January 2, 2019 Accepted: February 5, 2019 Letter pubs.acs.org/JPCL Cite This: J. Phys. Chem. Lett. 2019, 10, 780-785 © XXXX American Chemical Society 780 DOI: 10.1021/acs.jpclett.9b00009 J. Phys. Chem. Lett. 2019, 10, 780-785 Downloaded via KOREA INST SCIENCE AND TECHNOLOGY on February 13, 2019 at 23:19:01 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.