A Neural Network Approach for Predicting Microstructure Development in Cement

Although portland cement concrete is the most widely used construction material in the world, efficient simulations for predicting hydration and property development (hardening and aging) are yet to be developed and made widely available to researchers and engineers. Current state-of-theart simulations rely on computationally expensive models to generate spatially resolved time sequences of chemical and physical properties (microstructures) that allow engineers to predict permeability, strength, and durability. Improving simulation accuracy or efficiency would be a significant contribution. We describe a novel application of neural networks to simulate the microstructure of hydrating tricalcium silicate (the most abundant component of ordinary portland cement). Initial model predictions correlate well with benchmark models but require only a small fraction of the computation time.