Fundamental Factors Governing Stabilization of Janus 2D-Bulk Heterostructures with Machine Learning.

The more than 6000 2D materials predicted thus far provide a huge combinatorial space for forming functional heterostructures with bulk materials, with potential applications in nanoelectronics, sensing, and energy conversion. In this work, we investigate nearly 1000 heterostructures, the largest number of heterostructures thus far, of 2D Janus and bulk materials' surfaces using simulations and machine learning (ML) to deduce the structure-property relationships of the complex interfaces in such heterostructures. We first perform van der Waals-corrected density functional theory simulations using a high-throughput computational framework on 51 Janus 2D materials and 19 metallic, cubic phase, elemental bulk materials that exhibit low lattice mismatches and low coincident site lattices. The formation energies of the resultant 1147 Janus 2D-bulk heterostructures were analyzed, and 828 were found to be thermodynamically stable. ML models were trained on the computed data, and we found that they could predict the binding energy and -separation of 2D-bulk heterostructures with root mean squared errors (RMSEs) of 0.05 eV/atom and 0.14 Å, respectively. The feature importance of the models reveals that the properties of the bulk materials dominate the heterostructures' energies and interfacial structures heavily. These findings are in line with experimentally observed behavior of several well-known 2D materials-bulk systems. The data used within this paper are freely available in the 2D-Bulk Heterostructure Database (aiHD). The fundamental insights into 2D-bulk heterostructures and the predictive ML models developed in this work could accelerate the application of thousands of 2D-bulk heterostructures, thus stimulating research within a wide range of electronic, quantum computing, sensing, and energy applications.