Machine learning models for predicting ultimate bond strength of grouted sleeve connections

Grouted sleeve connection (GSC) is widely utilized in prefabricated concrete structures and accurately predicting its ultimate bond strength between rebar and grouting materials of GSC is crucial for ensuring structural safety. In this study, a parameter-sharing residual multi-fidelity neural network (PsRMFNN) is proposed to predict the ultimate bond strength of GSC. For this purpose, a database comprising 209 existing GSC tensile experimental data is established to train and evaluate the PsRMFNN. Comparative analyses are conducted with existing empirical equations, backpropagation neural networks, and residual multi-fidelity neural networks. The PsRMFNN exhibits superior performance, yielding R2 of 0.994, MAE of 6.17 kN, RMSE of 12.97 kN, and MAPE of 3.55% across all data. Furthermore, comparative evaluations of the three neural network-based models show that the PsRMFNN accelerates convergence and enhances prediction accuracy. Finally, based on the sensitivity analysis, predictive equations for bond strength are developed using the genetic algorithm and the genetic programming. The outcomes demonstrate that neural network-based models and newly proposed predictive equations can effectively forecast the ultimate bond strength between rebar and grouting materials of GSC.