Dresden 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

O: Fachverband Oberflächenphysik

O 50: Electronic structure theory II

O 50.5: Talk

Tuesday, March 10, 2026, 15:30–15:45, TRE/PHYS

Quantifying conductance variations in single molecule junctions using machine learning — •Hector Vazquez — Inst. of Physics, Czech Academy of Sciences

In single molecule circuits, conductance (the inverse of resistance) is strongly dependent on the junction geometry. Break-junction experiments are often carried out at room temperature, where many molecular conformations are sampled during the measurements. In contrast, the computational cost of DFT-NEGF calculations restricts them to only a few geometries.

We recently developed a computationally efficient method to calculate molecular conductance within DFT for thousands of geometries, based on small Au-molecule-Au clusters [1,2]. Their geometry is taken from MD simulations of the junction at room temperature. We can thus compute within DFT the conductance for tens of thousands of thermally-accessible molecular geometries.

We study typical conjugated and alkane molecules and interpret these large conductance datasets with machine learning methods including regression models, feature importance techniques, and SHAP analysis. Our work identifies which of the bond lengths, angles, or dihedral angles in the molecule, all of which are changing continuously and simultaneously, have a larger impact on conductance.

[1] H. Vazquez, J. Phys. Chem. Lett. 13 9326 (2022)

[2] E. Montes, W.Y. Rojas and H. Vázquez, J. Phys. Chem. C 129, 9947 (2025)

Keywords: Molecular electronics; Machine learning; Single molecule conductance; Structure-conductance relationships; Feature importance