Boosting current-induced molecular dynamics with machine-learning potential

Abstract

In a current-carrying single-molecular junction (SMJ), a hierarchy of hybrid energy transport processes takes place under a highly nonequilibrium situation, including energy transfer from electrons to molecular vibrations via electron-vibration interaction, energy redistribution within different vibrational modes via anharmonic coupling, and eventual energy transport to surrounding electrodes. A comprehensive understanding of such processes is a prerequisite for their potential applications as single-molecular devices. $Ab$ $initio$ current-induced molecular dynamics (MD) is an ideal approach to address this complicated problem. But the computational cost hinders its usage in systematic study of realistic SMJs. Here, we achieve orders of magnitude improvement in the speed of MD simulation by employing machine-learning potential with accuracy comparable to density functional theory. Using this approach, we show that SMJs with graphene electrodes generate order of magnitude less heating than those with gold electrodes. Our work illustrates the superior heat transport property of graphene as electrodes for SMJs, thanks to its better phonon spectral overlap with molecular vibrations.