Answers are grounded in this paper and cite their sources.
We present a machine-learning framework for diabatization based on the parallel transport of electronic states. Starting from a differential-geometric description of the Born–Oppenheimer approximation, we learn the generator of a continuous unitary transformation between adiabatic and quasi-diabatic electronic states. The generator is represented by a neural network and integrated along nuclear paths using a neural ordinary differential equation. Sparse non-adiabatic coupling data are used to train the local transport, while the more densely available adiabatic energies are included through an energy-weighted regularization term. This avoids fitting the adiabatic potential energy surfaces directly and preserves both orthonormality and the adiabatic energy spectrum by construction. The method is demonstrated for an analytically solvable two-state model with a conical intersection, the ionic–neutral avoided crossing in LiF, and a two-dimensional cut through the conical intersection of NO2. The examples show the construction of quasi-diabatic representations under explicitly specified geometric path conventions.
No takes yet. Share an insight, caveat, or question.
Hütter et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: