High-throughput exploration of stable semiconductors using deep learning and density functional theory

Abstract

Semiconductors can lead to new applications and technological innovations. In this work, we developed a computational pipeline to discover new semiconductors by combining deep learning and high-throughput first-principles calculations. We used a random strategy combined with group and graph theory to generate initial boron nitride polymorphs and developed a classifier based on graph convolutional neural network to screen semiconductors and study their stability. We found 26 new stable boron nitride polymorphs in Pc phase, of which 3 are direct bandgap semiconductors, and 10 are quasi-direct bandgap semiconductors. This discovery not only expands the library of known semiconductor materials but also provides potential candidates for high-performance electronic and optoelectronic devices, paving the way for future technological advancements.