Featurization strategies of supplementary cementitious materials for enhancing interpretability and precision in machine learning based concrete strength prediction

Abstract

Abstract This study investigates the impact of input variable processing methods for supplementary cementitious materials (SCMs) on the performance of machine learning (ML)-based concrete strength prediction models. SCMs possess unique pozzolanic and latent hydraulic mechanisms. However, it has not been clearly established whether it is more advantageous to treat variables individually to capture SCM characteristics or to aggregate them into a single variable to avoid the curse of dimensionality. This study compared and analyzed the predictive performance between strategies of individualizing versus combining SCM components using 13 databases. Furthermore, SHAP analysis was used to interpret the physical impact of each strategy on the internal decision-making mechanisms of ML models. The results showed that, in most cases, the individualized model recorded superior accuracy compared to the combined model. Particularly in environments where the strength contribution of each SCM component varied in direction, variable separation prevented the information cancel-out effect, allowing the model to capture physical causality more precisely. Conversely, in simple environments, where the influence of a specific variable such as age was overwhelming or the entire SCM exhibited a uniform trend in strength development, combining SCMs as a single variable could be advantageous for stable performance.