RFNET: Integrating Rich Features into Neural Networks for Band Gap Prediction of Perovskite Crystals

Perovskite crystals with simplicity in manufacturing and tuneable band gaps attracted wide attention in material science. Currently, the development of materials with specific band gaps remains difficult and consumes significant manufacturing resources. Therefore, demand keeps increasing in the context of material property prediction through machine learning in order to refine the discovery process. Herein, we proposed a novel model RFNET which integrated rich features into neural networks for predicting band gaps of Perovskite crystals. A virtual screen was studied to showcase the effectiveness of RFNET model in identifying the narrowest band gap perovskite crystals. We comprehensively compared the RFNET with nine other common machine learning and deep learning methods. The experimental result demonstrated that RFNET could reduce the number of candidate materials by 8% to 14%. We also offered a highly practical hands-on tutorial for material science researchers to reproduce the code of this work. Overall, this unprecedented model guaranteed the implications for enhancing virtual screening performance and minimizing workload.