Balancing predictive accuracy and engineering transparency: Interpretable additive machine learning for concrete compressive strength modeling

Modelling concrete compressive strength is a non-linear regression problem influenced by complex interactions between material composition and concrete curing age. The process of accurate prediction serves as a vital component which helps designers create materials and construction professionals perform quality control operations. Four regression models include Explainable Boosting Machine (EBM), Neural Network based on Generalized Additive Model with Structured Interaction (GAMI-Net) and Fast Interpretable Greedy-tree Sums (FIGS) and Rectified Linear Unit-based Neural Network (ReLU-NN) are employed as comparative models and evaluated using MSE, MAE and R-squared. The results indicate that EBM provides more accurate and stable predictive performance than the other three models in capturing non-linear patterns in the data. Model interpretation analysis further reveals that concrete age and cement content have dominant effects on compressive strength, while increased water content shows an opposite relationship. These findings demonstrate that interpretable machine learning models can achieve high predictive accuracy while maintaining model transparency in concrete compressive strength modeling.