Prediction model for compressive strength and drying shrinkage of alkali-activated materials: Evaluation of XGBoost and LightGBM

Alkali-activated materials (AAMs), featured by its cementless characteristic, received broad acceptance from numerous researchers these years. However, it still remains challenging to manufacture AAM mixtures with high mechanical performance and low drying shrinkage. A significant cause of the difficulty in controlling the mechanical properties of AAMs is the absence of an appropriate model for predicting compressive strength and drying shrinkage. In this study, a comparison study was conducted to predict the compressive strength and drying shrinkage of AAMs using an extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM) based on the key factors of the AAM mix design. Supervised machine learning algorithms were employed for the XGBoost and LightGBM methods to process the datasets through training, validation, modeling, and testing. In addition, the K-fold cross-validation method was adopted for the training dataset, in which the K values ranged from 2 to 10. The results showed that the R2 values of the best performed XGBoost and LightGBM models for strength prediction were 0.8787 and 0.8583, while for the drying shrinkage prediction R2 values of the XGBoost and LightGBM methods were 0.9906 and 0.8989, respectively. The above four models were all obtained with 10-fold cross validation. Moreover, for proving the applicability of the proposed models in the real construction work, a validation experiment for the proposed models was carried out in laboratory, and the model was able to estimate with an error of 15%. This research largely helps the decision-makers to properly use the XGBoost and LightGBM algorithms.