Universal materials model of deep-learning density functional theory Hamiltonian

Large materials models — analogous to large language models but for materials science — could revolutionize materials discovery, but building one that handles arbitrary compositions and structures is an open challenge. We demonstrated a concrete path forward: by constructing a diverse materials database and substantially improving the DeepH method, we trained a universal DeepH model that accurately predicts DFT Hamiltonians across diverse elemental compositions and crystal structures. We further showed how this universal model can be fine-tuned for specialized applications, laying the groundwork toward large materials models.