Abstract
•Light GBM and XGBoost prediction models for estimating RCPT of metakaolin-containing concrete.•Accuracy Analysis of the developed models using statistical evaluation.•Optimization of water-binder ratio and percentage of metakaolin using interpretable machine learning SHAP analysis.•Light GBM surpasses in accuracy compared to XGBoost.
This study investigates the non-linear capabilities of two machine learning prediction models, namely Light GBM and XGBoost, for predicting the values of Rapid Chloride Penetration Test (RCPT). Chloride penetration is one of the most significant issues affecting reinforced concrete (RC) structures, which necessitate frequent maintenance and repair. The mix design of concrete play a vital role in the formation of pore structure that is relatively more resistant to chloride attacks. For estimating the chloride resistance of concrete, 201 experimental records, incorporating aging of concrete, binder content, water-binder ratio, percentage of metakaolin, and content of fine and coarse aggregates as input variables. The models were trained using grid search optimization for tuning setting parameters to yield the best performance for the models. The performance of the models using statistical indices indicated LightGBM surpasses in prediction accuracy as compared to XGBoost model. The coefficient of determination (R2) values revealed 0.9738 and 0.9379 for LightGBM and XGBoost models, respectively. The minimum value of MAE was recorded for the training data of the LightGBM model equalling 172.7 C. The best fit model, i.e., the LightGBM model, was used for SHAP analysis to see the relative importance of contributing attributes and optimization of input variables. The SHAP analysis revealed fc’, aging, and W/B ratio as most significant in yielding RCPT, whereas individual interpretation of Shapley values showed that W/B ratio of 0.30 – 0.35 and 15% MK replacement highly resisted chloride penetration at higher compressive strength values.