The role of deep learning in reducing computational cost when simulating chloride ion attack on hydrated calcium silicate with molecular dynamics

Chloride anion attack is a major factor limiting the durability of concrete structures. To clarify the mechanisms by which chloride salts degrade concrete, nanoscale molecular dynamics (MD) simulations were used to study chloride attack on calcium silicate hydrate (CSH), the main component of cement. In MD simulations, a relaxation process is generally required to allow the system to reach equilibrium. However, relaxation is computationally expensive when performing MD simulations of large structural systems. This expense could potentially be avoided by using deep learning techniques. This paper describes the creation of a multi-fidelity physics-informed neural network model of a CSH gel pore containing an aqueous NaCl solution. The neural network's input variables are the ambient temperature and the NaCl concentration and its output variables are the system's energy, the Na-O radial distribution function, and the Na+ and Cl- ion density distributions. After training the model using the results of low-fidelity MD simulations without relaxation and a smaller number of high-fidelity simulations with relaxation, highly accurate outputs were obtained with prediction errors below 3%. Deep learning can thus greatly reduce the computational cost of MD studies of large and complex systems with no appreciable loss of accuracy.