摘要
Abstract Molecular electronics offers a route toward continued device miniaturization, but reliable control of charge transport at the molecular scale requires a detailed understanding of the metal–molecule interface. In particular, the role of electrode geometry in determining transport behaviour remains less systematically explored than molecular structure and contact chemistry. Here, we investigate the influence of atomic-scale electrode geometry on the structural, electronic, quantum-transport, thermoelectric, and noise properties of Au–pentane–Au molecular junctions using a combined density functional theory–non-equilibrium Green's function (DFT–NEGF) and machine learning framework. Six representative electrode configurations—Flat Au(111), Pyramidal, Single Apex, Multi-Atom Tip, Nanowire, and Rough Surface—were structurally optimized and systematically compared while maintaining the same pentane molecular backbone and Au–S contact chemistry. The calculations reveal pronounced geometry-dependent variations in Au–S bonding, coordination number, contact area, molecular tilt, charge transfer, electronic-state alignment, and molecule–electrode coupling. These structural and electronic changes directly control quantum transport, with the calculated conductance decreasing from 0.92 G₀ for Flat Au(111) to 0.56 G₀ for the Single Apex configuration. Transmission, current–voltage characteristics, charge redistribution, and interface dipole analysis consistently demonstrate that highly coordinated and extended contacts provide stronger electronic coupling and more efficient transmission pathways. Thermoelectric calculations reveal an opposing trend: lower-coordination contacts exhibit enhanced Seebeck coefficients because of stronger energy dependence of transmission near the Fermi level, whereas the Flat Au(111) junction provides the highest calculated power factor and ZT ≈ 0.68. Shot noise, Fano factor, low-frequency noise, and Hooge-parameter analysis further demonstrate increasing transport fluctuations as electrode coordination decreases. To accelerate transport prediction, a dataset of 1500 configurations were constructed from structural and electronic descriptors extracted from the DFT–NEGF calculations. Among the tested models, XGBoost achieved the highest predictive performance, with a testing R 2 of 0.988, RMSE of 0.010 G₀, and MAE of 0.007 G₀. Feature-importance analysis identified Au–S bond length, HOMO–LUMO gap, coupling strength, DOS at the Fermi level, and Au 5 d PDOS as the dominant predictors. The trained model successfully predicted conductance for previously unseen electrode configurations, demonstrating its potential for rapid screening of molecular junction structures.