Seonglae ChoFlow Matching domain-specifically extended/designed case. Uses equivariant prior + Hamiltonian-specific architecture. Density functional theory
Hamiltonian traditionally requires recalculation of Hamiltonian through SCF (Self-Consistent Field) iterations, resulting in high computational cost and slow performance. Previous attempts to directly predict Hamiltonian using ML mostly relied on simple regression → ignoring structural distribution/symmetry.
Learns trajectory from prior to target Hamiltonian distribution using CNF (Continuous Normalizing Flow) structure, training with vector field regression instead of simulation. Uses high-dimensional equivariant NN considering block-level symmetry of Hamiltonian and preserving rotational/translational symmetry (SE(3)-Equivariant Vector Field). Reflects structural constraints of Hamiltonian from initial distribution using Invariant Priors: GOE (Gaussian Orthogonal Ensemble), TE (Tensor Expansion) (group theory-based symmetry reflection). Fine-tuning for Energy Alignment by matching HOMO/LUMO, orbital energy derived from predicted Hamiltonian with actual DFT values.