How Molecular Packing Controls Electronic Structure in π-Stacked Radical Dimers: A DFT and Descriptor-Based Machine-Learning Study

Abstract π-Stacked radicals exhibit electronic properties that are highly sensitive to subtle changes in molecular packing, yet identifying which structural features control these interactions remains challenging. Here, we combine density functional theory calculations with chemically defined structural descriptors to quantify geometry−electronic structure relationships across 2,582 π-radical dimer configurations of five different polycyclic aromatic hydrocarbon (PAH) radical dimers. Extremely randomized trees and multilayer perceptron models predict fragment-orbital SOMO coupling, interaction energy, frontier-orbital separation, and the SOMO−SOMO interaction energy with test-set R2 values of 0.94−0.98 for radical families represented in training. Feature-importance analysis separates two structural regimes: global π-surface descriptors are most strongly associated with frontier-orbital variation and contribute to the interaction-energy scale, whereas local stacking registry and short C···C contact networks are most strongly associated with radical−radical coupling and corrected spin-dependent interaction energy. Comparison with a baseline that uses no structural descriptors shows that the frontier-orbital separation is largely set by the identity of the parent radical, whereas the two pancake-bonding properties are predicted from packing geometry alone and transfer to radical families absent from training. Y-randomization and geometry-clustered validation support the robustness of these relationships. This framework provides an interpretable, geometry-based route for evaluating and prioritizing packing arrangements in π-stacked radical dimers.