Regensburg 2025 – scientific programme

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

O: Fachverband Oberflächenphysik

O 92: Electronic Structure Theory

O 92.3: Talk

Thursday, March 20, 2025, 15:30–15:45, H25

Learning conductance of aromatic and antiaromatic molecular junctions — •Mohammad Ali Mohammadi Keshtan and Hector Vazquez — Inst. of Physics, Czech Academy of Sciences

Single-molecule junction conductance depends on molecular conformation. The standard computational method to study the conductance of a molecular junction is DFT-NEGF. However, DFT-NEGF is computationally expensive, so finding a fast and accurate method to compute the conductance of a large numbers of structures is vital.

Here, we use machine learning (ML) methods, including kernel ridge regression and Gaussian process regression, to overcome these limitations. To train the regression models, we first generate thousands of junction geometries using classical molecular dynamics at room temperature. For each geometry we then compute conductance using a computationally efficient approximation which considers a Au-molecule-Au complex [1], and build SOAP and Coulomb matrix descriptors.

We study a pair of aromatic/antiaromatic porphyrin-like molecules [2], whose large size further hampers the use of DFT-NEGF. We explore the performance of the different ML models, compare the results with DFT-NEGF, and discuss the relative importance of the different descriptors. Our work demonstrates how ML models can be efficiently trained and used to compute single molecule junction conductance.

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

[2] S. Fujii et al., Nat. Commun. 2017, 8, 15984

Keywords: Machine learning; Single molecule junction; Conductance; DFT-NEGF