Machine-learning prediction of the formation of atomic gold wires by mechanically controlled break junctions

One of the challenging issues in the formation of atomic wires in break-junction experiments is the formation of stable monoatomic chains of reasonable length. To address this issue, in this study, we present a combination of unsupervised and supervised machine learning models trained on the experimentally measured conductance traces, with a goal to develop a microscopic understanding. Applying a machine learning model to two independent data sets from two different samples containing 72 000 and 90 000 conductance-displacement traces of single-atomic junctions, respectively, we first obtain the optimum conditions of bias and the stretching rate for the formation of chains of length > 4 angstrom. A deep learning method is subsequently applied for the classification of individual breaking traces, leading to the identification of trace features related to long-chain formation. Further investigation by ab initio molecular dynamics simulations provides a molecular-level understanding of the problem.