A machine learning strategy for the physics-informed design of nanomaterials using ChIMES

Nanoparticles have an intrinsic ability to self-assemble into materials crucial to various industries spanning catalysis, drug delivery, optics, and environmental remediation. Simulation is a critical tool for exploring and designing new self-assembling nanomaterials, yet developing suitable models can be cumbersome due to the need for a balance of computational efficiency (i.e., to enable many-particle simulations) and accuracy. In this presentation, we explore the application of ChIMES, a generalized many-body machine-learned interatomic model (ML-IAM) and artificial intelligence-driven parameterization capability, to the modeling of nanoparticle systems. We note that ChIMES was originally developed and optimized for modeling condensed phase reacting systems at atomistic resolution. Here, we apply it to the development of models coarse-grained at the resolution of individual nanoparticles, which should, in principle, overcome limitations of commonly applied molecular mechanics functional forms, opening the door to modeling more complex nanoparticle systems (e.g., for which greater-than-2-body interactions cannot be neglected). We explore model sensitivity with respect to the choice of hyperparameters and discuss findings within the context of previously established “best practices” for atomistically-resolved ChIMES models.