Predicting mechanical properties of carbon nanotube-reinforced cementitious nanocomposites using interpretable ensemble learning models. (7th November 2022)

• Record Type: Journal Article
• Authors: Adel, Hossein; Palizban, Seyed Mohammad Mahdi; Sharifi, Seyed Sina; Ilchi Ghazaan, Majid; Habibnejad Korayem, Asghar
• Is Part Of: Construction & building materials. Volume 354(2022)
• Publication Date: 2022-11-07
• Subjects: Carbon nanotube (CNT) -- Cementitious nanocomposite -- Machine learning -- Ensemble learning -- Sensitivity analysis -- SHAP; Building materials -- Periodicals; 624.18
• DOI: 10.1016/j.conbuildmat.2022.129209
• Languages: English
• ISSNs: 0950-0618

Abstract

Highlights: Two comprehensive data sets are collected from reports that are published after 2016 for flexural and compressive strength. Four algorithms based on ensemble learning are used to predict the mechanical strength of CNT-reinforced composites. SHAP, Gini importance, F-score, and permutation importance are employed to quantify factors' influence. XGBoost surpasses others in accuracy and variables relationship detection. Abstract: Compressive and flexural strength are important characteristics that indicate the efficiency of utilizing carbon nanotube (CNT) in cementitious nanocomposites. Preparing numerous test samples and conducting mechanical tests at different ages is costly and time-consuming, so a predictive model appears essential. Due to the complexity of recent building materials, experimental and statistical models do not function well for such complex materials, so machine learning techniques were adopted in this study. From the results of experiments performed in the literature, separate comprehensive data sets for compressive and flexural strength were collected, and water to cement ratio (W/C), CNT type, CNT content, CNT length, CNT diameter, sand to cement ratio for mortars (S/C), surfactant type, dispersion method, curing days (age) and compressive or flexural strengths of the control sample (C0 or F0) considered as input variables. Four ensemble learning-based algorithms, including random forest, AdaBoost, Gradient Boost, and XGBoost, were developed alongside a standalone model, the decision tree. To assess the sensitivity of input variables, four methodologies, including the Gini importance, permutation importance, F score, and mean of relative Shapely additive explanation (SHAP) values, were obtained, and compressive or flexural strengths of the control sample and CNT content identified overall as the most influential variables. The results revealed that XGBoost provides more reliable and accurate results for compressive and flexural characteristics and better maps the relationships between input variables. In addition, SHAP analysis was performed to explain the contribution of each input according to its value on the output of XGBoost.