Optimisation of silane chemistry with machine learning and molecular dynamics

Among the numerous silane chemistries in commercial use and literature, only a few are well-parameterised and understood. This study presents a systematic approach using molecular dynamics (MD) and machine learning to rank and identify molecules with optimal mechanical strengths. MD simulations were conducted on silane molecules of varying functionalities and molecular weights to assess peak force performance in Mode I. The systems comprised a single silane molecule bonded to a glass surface through one or two bonds in either extended or relaxed configurations. Based on bond failure mechanisms, silane molecules were classified into three failure categories: (1) Silicon-Oxygen/Silicon-Carbon scission, (2) Nitrogen-Carbon scission, and (3) Oxygen-Carbon scission. In relaxed configurations, non-bonded interactions between silane molecules and the glass surface contributed significantly to energy absorption, with about 25% of total energy utilised to overcome these interactions. A machine learning approach was employed to predict peak forces for over 84,000 molecules by leveraging parameters from MD simulations. A Variational Autoencoder was used to reduce the molecular structure information to a low-dimensional, physically-informed space, followed by a quantile random forest regression model to predict the peak force performance for the broader set of silane molecules.