Machine learning the local electronic density of states

Electronic density of states (DOS) plays a crucial role in determining and understanding materials properties. We investigate the machine learnability of additive atomic contributions to electronic DOS, focusing on atom-projected DOS rather than structural DOS. This approach for structure-property mapping is both scalable and transferable, and achieves high prediction accuracy for pure and compound silicon and carbon structures of various sizes and configurations. Furthermore, we demonstrate the generalizability of this model to complex Sn-S-Se compound structures. Utilizing locally trained DOS is shown to significantly enhance the accuracy of predicting material properties, including band energy, Fermi energy, 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 examination of electronic structures in materials without the need for computationally expensive ab initio calculations.