Prediction & optimization of alkali-activated concrete based on the random forest machine learning algorithm

Abstract

Alkali-activated concrete (AAC) is regarded as a promising alternative construction material to reduce the CO₂ emission induced by Portland cement (PC) concrete. Due to the diversity in raw materials and complexity of reaction mechanisms, a commonly applied design code is still absent to date. This study attempts to directly correlate the AAC mix design parameters to their performances through an artificial intelligence approach. To be specific, 145 fresh property data and 193 mechanical strength data were collected from laboratory tests on 52 AAC mixtures, which were used as inputs for the machine learning algorithm. Five independent random forest (RF) models were established, which are able to predict fresh and hardened properties (in terms of compressive strength, slump values, static/dynamic yield stress, and plastic viscosity) of AAC with equivalent accuracy reported in the literature. Moreover, an inverse optimization was performed on the RF model obtained to reduce the sodium silicate dosages, which may further mitigate the environmental impact of producing AAC. The present RF model gives practical information on AAC mix design cases.