Understanding domain reconstruction of twisted transition metal dichalcogenide bilayers through machine learned interatomic potentials

Abstract In the study of twisted bilayers of two-dimensional materials, a detailed picture of the relaxations and layer-corrugations that occur due to interlayer interaction is crucial for predicting how their electronic and optical properties depend on twist angle and the resulting large-scale moiré pattern. As the relative twist angle between the layers approaches 0 ∘ , referred to as parallel (P) stacking, or 60 ∘ , referred to as antiparallel (AP) stacking, reconstructions occur to maximize the area of low-energy stacking domains, with a lattice of nodes of high-energy stacking connected by domain walls (DWs). We show that machine learned interatomic potentials can provide sufficiently precise energetics of stacking, strain, shear and varying interlayer distances to be used in place of the corresponding ground-truth vdW-corrected density functional theory for systems dramatically larger than those that can be treated with ab initio methods. We predict, explain, and quantify the domain reconstruction patterns for all like-chalcogen combinations of the transition metal dichalcogenides MoS 2 , MoSe 2 , WS 2 and WSe 2 down to twist angles approaching 1 ∘ . We demonstrate effects including triangular and kagome-like patterns in low-twist P and AP systems, respectively, and twirling around domain nodes in heterobilayers. For homobilayers, we present and parameterize an energy model that decomposes the total energy into contributions from constituent structural units. We further provide an experimental comparison for the MoS 2 homobilayer, which shows strong agreement with observed reconstruction patterns.