Machine learning band gap predictions: linking quasiparticle self-consistent GW and LDA-derived partial density of states

Abstract Accurate calculation of band gaps for given crystal structures is highly desirable. However, conventional first-principles calculations based on density functional theory within the local density approximation (LDA) fail to predict band gaps accurately. To address this issue, the quasiparticle self-consistent GW (QSGW) method is often used, as it is one of the most reliable theoretical approaches for predicting band gaps. Despite its accuracy, QSGW requires significant computational resources. To overcome this limitation, we propose to combine QSGW with machine learning. In this study, we applied QSGW to 1516 materials from the Materials Project [ https://materialsproject.org/ ] and used machine learning to predict QSGW band gaps as a function of the partial density of states (PDOS) in LDA. By treating the PDOS as multi-channel spectral image data, our convolutional neural network model effectively extracts feature vectors representing the electronic structure, achieving higher accuracy than conventional descriptor-based methods. Our results demonstrate that the proposed model significantly outperforms linear regression approaches with linearly independent descriptor generation [ https://github.com/Hitoshi-FUJII/LIDG ]. This model is a prototype for predicting the properties of materials based on PDOS.