Exploring machine learning to predict the pore solution composition of hardened cementitious systems

Abstract

This study explores machine learning (ML) algorithms to predict the pore solution composition of hardened cementitious systems produced with Portland cement (PC) and supplementary cementitious materials (SCM). Literature data on pore solution compositions for different cementitious systems was collected and compiled in a comprehensive database containing >300 entries with >80 features. Improved decision tree regressors were applied to the database. It was found that the trained ML models were capable of predicting OH−, Na+, and K+ concentrations reliably (75–90 % of predicted systems within 25 % relative error). Ca2+ and sulfur species had lower prediction accuracy. The silica content of SCM, the alkalis content of PC, and the SCM replacement level were identified as important features in determining the ion concentrations. When applied to this database, ML performed better than conventional, theory-based prediction models. Thus, ML models are a promising, complementary technique to determine pore solution compositions.