Lateral heterostructures of graphene and hexagonal boron nitride (h-BN) exhibit promising properties that are highly sensitive to the patterning of alternating domains. However, the vast design space of domain arrangements has hindered systematic mapping of their properties. Herein, we develop a neuroevolution machine learning interatomic potential (MLIP) to efficiently and accurately predict the piezopotential properties of h-BN/graphene heterostructures. Using the trained MLIP, high-throughput molecular dynamics simulations show that the mechanical, piezoelectric, and dielectric responses of h-BN/graphene heterostructures can be widely tuned by varying the concentration and arrangement of graphene domains, rendering their piezopotentials highly designable. This tunability of piezopotential is primarily ascribed to the polarization-gradient and flexoelectric effects, majorly governed by two topological descriptors: the coefficient of variation and a symmetry index. To target the specified piezopotential, we further introduce a machine learning-based inverse-design method. This study not only expands the understanding of piezoelectric properties of h-BN/graphene lateral heterostructures but also provides a strategy to design two-dimensional lateral heterostructures with tailored electromechanical properties for high-performance flexible piezoelectric devices.
Wang et al. (Thu,) studied this question.