Prediction model of long-term tensile strength of glass fiber reinforced polymer bars exposed to alkaline solution based on Bayesian optimized artificial neural network

• Published in: Construction & building materials Vol. 400; p. 132885
• Main Authors: Dong, Shaoce, Wu, Xu, Qi, Xiao, Affolter, Christian, Terrasi, Giovanni P., Xian, Guijun
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 12-10-2023
• Subjects: Alkaline ageing; Artificial neural network; Bayesian optimization; Creep strength in alkaline media; Glass fiber reinforced polymer bars; Long-term tensile strength; Long-term tensile strength; Alkaline ageing; Artificial neural network; Bayesian optimization; Glass fiber reinforced polymer bars; Creep strength in alkaline media
• ISSN: 0950-0618; 1879-0526
• DOI: 10.1016/j.conbuildmat.2023.132885

•Preassessment of the durability dataset of thermoset matrix based GFRP bars.•An ANN model assisted by BO search was built with excellent prediction ability.•Interpretation of the optimal ANN model by PDP plot and SHAP values calculation.•Limitations of machine learning models in predicting the retentions of GFRP bars.
The present study presents a model for predicting the long-term tensile strength of glass reinforced thermoset polymer (GFRP) bars based on the artificial neural network (ANN) algorithm, for bars which were exposed to alkaline solutions with different temperatures, different chloride ions concentration, and different applied stress levels. 272 groups of durability data of GFRP bars were collected from the literature. The correlation coefficients between each input feature and the target variable were calculated to analyze their relationships. The principal component analysis (PCA) method was performed to identify possible opportunities to reduce the input dimensions. Bayesian optimized search and randomized search were used and compared to tune and optimize the hyperparameters of the ANN model. Once the optimal ANN model was built, permutation feature importance, partial dependence plot (PDP), and SHapley Additive exPlanations (SHAP) value methods were used to interpret the optimal model. Results indicate that only the ultimate tensile strength has a high potential of being linearly changing with the long-term tensile strength of GFRP bars. PCA results show that the input dimensions cannot be reduced if 99% explained variance needs to be retained. The optimal ANN model trained through the Bayesian optimized search method can achieve a determination coefficient R2 and a mean absolute percentage error of 0.941 and 7.11%. Permutation feature importance and SHAP importance both ranked the initial ultimate tensile strength of GFRP bars as the most important input feature, followed by temperature, ageing time, diameter, applied stress level, pH value, and chloride ions concentration based on their SHAP values. On average, the PDP result of the ageing time tended to show that there will be a certain residual tensile strength of GFRP bars after ageing for a long period.