Abstract

Semiconductor materials for energy storage are the core and foundation of modern information society and play important roles in photovoltaic system, integrated circuit, spacecraft technology, lighting applications, and other fields. Unfortunately, due to the long experiment period and high calculation cost, the high-precision band gap (the basic characteristic parameter) of semiconductor is difficult to obtain, which hinders the development of new semiconductor materials. Since the traditional Perdew–Burke–Ernzerhof (PBE) functional not only requires a long calculation time, but also significantly underestimates the band gap, we developed a deep learning model that can predict the more precise Heyd–Scuseria–Ernzerhof (HSE06) band gaps in milliseconds for 1,503 binary metallic oxides, nitrides, and sulfides, with a mean absolute error (MAE) of 0.35 eV, a mean squared error (MSE) of 0.21 eV, and a coefficient of determination (R 2 ) of 0.98. Based on transfer learning, only < 5% of the data set (64 structures) was required to train the model and predict the band gaps of the remaining 1,439 structures. From the 1,503 candidate materials, we quickly identified 75 carrier transport materials, 33 electrode and electrocatalytic materials, 299 power switching materials, and 114 sensing materials. This work is the first to demonstrate the feasibility of transfer learning in band gap prediction, from the low-level PBE to the high-level HSE06 calculation, with a computation speed at least 10 4 times faster than the ab initio calculation. The proposed method could be further expanded to incorporate entire organic/inorganic crystal materials databases ( > 10 6 crystals), which is of great significance for the screening and discovery of new semiconductor energy storage materials. 1.

Citations

Authors (7)