Fundamental Band Gap and Alignment of Two-Dimensional Semiconductors Explored by Machine Learning

TL;DR

This paper combines density functional theory and machine learning to predict fundamental band gaps and alignments of over 400 2D semiconductors, enabling efficient materials design for next-generation electronics.

Contribution

It develops machine learning models, especially Support Vector Regression, to accurately predict electronic properties of 2D semiconductors, facilitating targeted materials design.

Findings

SVR achieves RMSE < 0.15 eV for band gaps and alignments

Machine learning models effectively screen 2D materials for desired electronic properties

Family of 2D semiconductors shows wide band gap distribution from 0 to 8 eV

Abstract

Two-dimensional (2D) semiconductors isoelectronic to phosphorene has been drawing much attention recently due to their promising applications for next-generation (opt)electronics. This family of 2D materials contains more than 400 members, including (a) elemental group-V materials, (b) binary III-VII and IV-VI compounds, (c) ternary III-VI-VII and IV-V-VII compounds, making materials design with targeted functionality unprecedentedly rich and extremely challenging. To shed light on rational functionality design with this family of materials, we systemically explore their fundamental band gaps and alignments using hybrid density functional theory (DFT) in combination with machine learning. First, GGA-PBE and HSE calculations are performed as a reference. We find this family of materials share similar crystalline structures, but possess largely distributed band-gap values ranging approximately from 0 to 8 eV. Then, we apply machine learning methods, including Linear Regression (LR), Random Forest Regression (RFR), and Support Vector Machine Regression (SVR), to build models for prediction of electronic properties. Among these models, SVR is found to have the best performance, yielding the root mean square error (RMSE) less than 0.15 eV for predicted band gaps, VBMs, and CBMs when both PBE results and elemental information are used as features. Thus, we demonstrate machine learning models are universally suitable for screening 2D isoelectronic systems with targeted functionality and especially valuable for the design of alloys and heterogeneous systems.