Machine-Learning Assisted High-Throughput Computational Identification of Mixed-Dimensional Materials

The nano revolution started with the discovery of graphene and has led to a great deal of research into low-dimensional systems. These systems range from 0D nanoparticles to 1D nanowires and nanoribbons to 2D monolayers, all of which experience quantum confinement that can give rise to unique properties. More recently, mixed dimensional materials have received increasing attention due to the potential for synergistic and emergent properties in combinations of 2D, 1D, and 0D systems. These investigations have used intentionally synthesized mixed dimensional systems, but the question of what mixed dimensional materials occur in nature has yet to be tackled. In this work, we search the MaterialsProject database for crystals that contain sub-networks of atoms with differing dimensionalities (e.g., 1D nanowires between 2D layers). We perform this search with the Topological Scaling Algorithm, using atomic radii as a first pass and the CrystallNN algorithm as a second. We follow by using the Electron Localization Function to train an ensemble of neural network classifiers to distinguish bonds with a set of well-chosen materials examples. Subsequently, we apply the trained models to classify bond types and identify candidate mixed-dimensional heterostructures from a broad range of materials systems as a third pass. From these searches, we identify several mixed-dimensional crystal structures, which we then relax using Density Functional Theory while accounting for van der Waals forces. Finally, we calculate various properties in these crystals, including the electronic structure and thermodynamic stability.