UniTB framework for universal equivariant Hamiltonian learning

UniTB introduces an E(3) equivariant Hamiltonian representation that integrates density-functional tight-binding theory with deep learning. The model uses a mixture-of-experts architecture and a charge-equilibration head to generalize across different chemistries and large supercells while suppressing spurious states through projection bounds.