Machine-Learning-Driven Molecular Dynamics Unravels Stereoelectronic Switching in Statistical Ensembles of Single-Molecule Junctions.

Break-junction measurements provide a direct probe of charge transport through single-molecule junctions, but the junction is created and elongated under mechanical loading, leading to variable junction geometries and broad conductance distributions. Accordingly, although conductance histograms are reproducible over thousands of traces, the underlying junction conformations remain poorly defined. To address this problem, we combine machine learning force field (MLFF) with molecular dynamics (MD) to study the Au-oligosilane-Au molecular junctions (with methylthiomethyl [-CH2SMe] as linking groups) with stereoelectronic switching effects. In contrast to previous studies that proposed conformational changes only at the molecule-electrode interfaces, our simulations show that stretching the junction induces dihedral angle variations both within the molecular backbone and at the interfaces. Using a large ensemble of MD snapshots with computed conductance, we build a statistical configuration-conductance relationship. Principal component analysis (PCA) and correlation analysis identify two classes of dihedral coordinates that elucidate most of the variance in conductance. These key torsions tune the symmetry and delocalization of σ-type molecular electronic states and the ensuing couplings of these states with the two gold electrodes, producing up to 2 orders of magnitude changes in conductance. This work not only clarifies the complex evolution of junction structures at the atomic scale and the resulting stereoelectronic effects, but also promotes the mapping and understanding of the structure-property relationship in molecular electronic devices.