Accurate prediction of band gap of two-dimensional monolayer materials via transfer learning

Band gap is a crucial factor for the design and discovery of novel functional materials with desired properties. In principle, the band gap can be accurately predicted by using first-principles calculations with quasiparticle self-energy corrections, which is however very time-consuming and thus limited to small systems. In this work, using a pre-trained neural network architecture for the band gap calculated with standard Perdew-Burke-Ernzerhof (PBE) functional, we propose a transfer learning (TL) model to readily and accurately predict the band gap of any monolayer materials, where a small set of GW-calculated gaps is used as training data. Compared with general machine learning algorithms, the TL-driven model shows superior predictive performance, as manifested by improved Pearson correlation coefficient (reduced mean absolute error) from 71% to 97% (0.55 to 0.27) between the real and predicted band gaps. Importantly, the established TL model with good interpretability is leveraged to predict the GW gaps of 2915 monolayer systems that are retrieved from the Computational 2D Materials Database (C2DB) with non-zero PBE gaps, providing a rich sample space for exploring high performance functional materials with suitable band gap.