Machine learning to predict properties of fresh and hardened alkali-activated concrete

Abstract

Alkali-activated concrete (AAC) is widely considered to be a sustainable alternative to Portland cement concrete. However, on account of extensive heterogeneity in composition of the aluminosilicates, coupled with the failure of classical materials science approaches to unravel the underlying composition-property linkages, reliable prediction of AAC's properties has remained infeasible. This paper presents a random forest (RF) model to predict two properties of fly ash-based AACs that are important from compliance standpoint – slump flow; and compressive strength – in relation to physiochemical attributes, curing conditions, and mixing procedures of the concretes. Results show that the RF model – once meticulously trained, and after its hyperparameters are rigorously optimized – is able to produce high fidelity predictions of both properties of new AACs. The model is also used to quantitatively assess the influence of physiochemical attributes and process parameters on the AAC's properties. Outcomes of this work present a pathway to optimization of AACs' properties.