Interpretable Ensemble-Machine-Learning models for predicting creep behavior of concrete

This study aims to provide an efficient and accurate machine learning (ML) approach for predicting the creep behavior of concrete. Three ensemble machine learning (EML) models are selected in this study: Random Forest (RF), Extreme Gradient Boosting Machine (XGBoost) and Light Gradient Boosting Machine (LGBM). Firstly, the creep data in Northwestern University (NU) database is preprocessed by a prebuilt XGBoost model and then split into a training set and a testing set. Then, by Bayesian Optimization and 5-fold cross validation, the 3 EML models are tuned to achieve high accuracy (R2 = 0.953, 0.947 and 0.946 for LGBM, XGBoost and RF, respectively). In the testing set, the EML models show significantly higher accuracy than the equation proposed by the fib Model Code 2010 (R2 = 0.377). Finally, the SHapley Additive exPlanations (SHAP), based on the cooperative game theories, are calculated to interpretate the predictions of the EML model. Five most influential parameters for concrete creep compliance are identified by the SHAP values of EML models as follows: time since loading, compressive strength, age when loads are applied, relative humidity during the test and temperature during the test. The patterns captured by the three EML models are consistent with theoretical understanding of factors that influence concrete creep, which proves that the proposed EML models show reasonable predictions.

Automated summary

This study aims to predict the creep behavior of concrete using machine learning techniques, specifically EML models such as RF, XGBoost, and LGBM, with high accuracy achieved through Bayesian Optimization and cross-validation. The SHAP values highlight the significant influence of parameters like time, compressive strength, age, humidity, and temperature on concrete creep compliance, indicating the reasonableness of the EML model predictions.