Locally Machine-Learnability of Density of Electronic States
Abstract
The electronic density of states (DOS) plays a crucial role in determining the properties of materials. In this study, we investigate the machine learnability of additive atomic contributions to the electronic DOS. Our approach focuses on atom-projected DOS rather than structural DOS. This method for structure-property mapping is both scalable and transferable, achieving high prediction accuracy for pure and compound silicon and carbon structures of varying sizes and configurations. Furthermore, we demonstrate the effectiveness of our method on the complex Sn-S-Se compound structures. By employing locally trained DOS, we significantly enhance the accuracy in predicting secondary material properties, such as band energy, Fermi level, heat capacity, and magnetic susceptibility. Our findings indicate that directly learning atomic DOS, rather than structural DOS, improves the efficiency, accuracy, and interpretability of machine learning in structure-property mapping. This streamlined approach reduces computational complexity, paving the way for the examination of electronic structures in materials without the need for computationally expensive ab initio calculations.