Predicting the effect of SCMs on ASR in the accelerated mortar bar test with artificial neural networks

Supplementary cementitious materials (SCMs) are an efficient way to both mitigate ASR and reduce the carbon footprint of concrete. Identifying possible new materials and assessing their suitability regarding ASR have become major issues, since the amount of traditional SCMs such as fly ash is declining. Accelerated mortar bar test has been widely used to test different materials with various replacement levels, and a substantial amount of data is available in the literature. This study explores the possibility of using artificial neural networks to analyse this large dataset. Attention is drawn on the relationship between the chemical composition of the binder and the reduction in expansion brought by the addition of SCMs, compared to the reference mixes with cement only. Using a baseline case with only the CaO and SiO₂ contents of the binder as inputs, one can see that the individual addition of other compounds (Al₂O₃, Fe₂O₃, MgO, SO₃ and Na₂O_eq) does improve the neural network performance. Combining them altogether improves the performance even further, although the assumption of independence of inputs may no longer be valid. After training, the artificial neural network is able to predict with a relatively good accuracy the SCM effect: the reduction in expansion is successfully predicted within ± 20 percentage points for more than 90 % of the dataset. However, uncertainties remain on the quantitative effect of each oxide, which could be investigated further by performing other types of regression on the same dataset. Besides, increasing the dataset size to fully exploit the potential of artificial neural networks and investigating methods to shed light on the input-output relationship are also promising leads to strengthen the analysis.