GNN- ω PBE: Range-Separation Parameters Learned from Three-Dimensional Structures
TL;DR
This study advances machine learned density functional development from molecule-specific to geometry-specific modeling of conformational effects in electronic structures by introducing GNN-ωPBE, a second machine learned RSH functional that generates geometry-specific ω values directly from atomic numbers and three-dimensional atomic coordinates.
Abstract
Range-separated hybrid (RSH) functionals improve density functional theory (DFT) for modeling organic semiconducting molecules by reducing self-interaction and delocalization errors, but their accuracy depends critically on the range-separation parameter (ω). Although our previously developed ML-ωPBE functional generates molecule-specific ω values from topology-based descriptors, its applicability is limited for flexible molecules with rich conformational diversity. We introduce GNN-ωPBE, a second machine learned RSH functional that generates geometry-specific ω values directly from atomic numbers and three-dimensional (3D) atomic coordinates. Built on a graph neural network (GNN) representation, GNN-ωPBE encodes interatomic interactions through internal coordinates and distinguishes different conformers of the same molecule, enabling it to capture geometry-dependent variations inaccessible to topology-based molecular descriptors. Trained on a conformer-resolved dataset, GNN-ωPBE reproduces the benchmark OT-ωPBE functional with a mean absolute error (MAE) of approximately 0.0072 a0–1. Our study advances machine learned density functional development from molecule-specific to geometry-specific modeling of conformational effects in electronic structures.
