Combining Dataset Curation and Many-Body Calculations for describing Structurally Complex 2D and Quantum Materials

There is considerable excitement about the family of van-der-Waals-bonded materials, such as transition metal dichalcogenides (TMDs). They not only display unique electronic and optical properties when thinned down to a monolayer, but also allow for a radical new approach to making materials by judiciously stacking individual layers. However, there is a large phase space of possible layered materials when one considers the various possible interfacial stacking angles and layer chemical compositions, making a brute-force exploration of the problem via density-functional theory (DFT) unfeasible. We will discuss how to address these challenges with machine-learning- (ML-) parametrized classical force fields that capture the atomistic reconstruction in chemically nontrivial materials. In particular, we show how careful dataset generation and augmentation techniques can reduce the computational effort to train accurate ML force fields and yield structural information for chemically and symmetrically dissimilar materials. In parallel, we will comment on how similar techniques can be utilized to study the structural and dynamical properties of photoexcited materials. By combining accurate excited-state calculations based on the first-principles GW plus Bethe-Salpeter equation with machine-learning force fields, we can describe the dynamics of atoms after a material is optically excited, paving the way for the description of processes ranging from exciton-polaron formation to photochemical processes in structurally complex heterogeneous systems.