Exploring interpretable ensemble learning to predict mechanical strength and thermal conductivity of aerogel-incorporated concrete

• Published in: Construction & building materials 2023-08, Vol.392, p.131781, Article 131781
• Main Authors: Han, Fenglei, Lv, Yang, Liu, Yan, Zhang, Xuefu, Yu, Wenbing, Cheng, Chongsheng, Yang, Wei
• Format: Article
• Language: English
• Subjects: Aerogel-incorporated concrete; Ensemble learning; Mechanical strength; Performance prediction; SHAP; Thermal conductivity
• ISSN: 0950-0618; 1879-0526
• DOI: 10.1016/j.conbuildmat.2023.131781

•The compressive strength and thermal conductivity of aerogel-incorporated concrete were predicted successfully by ensemble learning.•The comparisons between the performance of proposed prediction models were performed.•The importance analysis of input variables using interpretable machine learning SHAP analysis.•Water-binder ratio and aerogel replacement rate were shown to have significant effect. As a new type of cementitious composite material with low carbon and environmental protection, the mechanical strength and thermal conductivity of aerogel-incorporated concrete (AIC) are important performance indicators. However, due to the complexity of its impacting components, it is challenging to forecast material performance only from the mix proportion. In this paper, 660 sets of test data obtained by laboratory investigation were adopted to explore a prediction method for mechanical strength and thermal conductivity of AIC based on interpretable ensemble learning. Important variables such as water-binder ratio, aerogel replacement rate, silica fume replacement rate, age, and dry/saturated state were selected as input parameters. The ensemble learning model was compared with four traditional machine learning algorithms. Cross-validation and grid search were applied to the hyperparameter optimization. The results showed that integrated learning was superior to traditional machine learning methods, and the MAPE of compressive strength, flexural strength and thermal conductivity decreased by 69.8%, 63.6% and 53.7% on average, respectively. In ensemble learning, the prediction accuracy of the Boosting algorithm was better than that of the Bagging algorithm, and the R2 of all prediction models was higher than 0.97. The LightGBM was the best in the prediction of compressive strength and thermal conductivity, while the XGBoost was in the prediction of flexural strength. The most important factor affecting mechanical strength of AIC was water-binder ratio and the replacement rate of aerogel, and the SHAP values could all reach above 20.0 MPa. For the thermal conductivity, only the replacement rate of aerogel had an obvious effect, and the SHAP value could achieve more than 1.0 W/m∙K.