Machine Learning for Prediction of Tribological Properties of Interfaces Between Single Crystals

The nanoscale interfacial structure between two single crystal surfaces can greatly affect friction and wear on macro-scales. For instance, the tribological behavior of bilayer graphene is controlled by the relative orientation between individual graphene layers, with select magic angles leading to superlubricity and near-zero coefficient of friction. However, there is no scheme to predict the tribological property of interfaces formed between two single crystals based on the parameters of the individual surfaces.

In this study, we investigate tribological properties of interfaces between two model single crystals to quantify the influence of crystal structures (BCC/FCC/HCP), crystallographic orientations of surfaces (hkl), twist angle (φ), surface energies (ε), and surface reconstruction on frictional sliding at their interface. The tribological behavior in interfacial configurations is assessed by calculating the potential energy surfaces, and energy profiles along the most facile sliding direction. Machine learning methods, trained on interfacial structures, featurized as crystal graphs, or potential energy surface images, are used to predict tribological properties, including energy barriers along easy sliding directions. Computed attention weights are used to identify the most relevant parameters that govern tribological properties and thus identify design rules for superlubricious interfaces. The effect of intracrystal and interfacial bond strengths on the computed tribological properties is also investigated in order to generalize the design rules for homo- and hetero-interfaces.