Bayesian optimization for selecting efficient machine learning regressors to determine bond-slip model of FRP-to-concrete interface

Abstract

Interfacial bond behaviour among FRP and concrete is a crucial aspect in determining the strengthening performance of FRP. A data-driven approach is proposed to automatically identify the key parameters to quantify the peak interfacial bond stress (τm) with the corresponding interfacial fracture energy (Gf), and consequently the bond-slip (τ-s) model can be determined as well. Hyperparameter optimization for selecting the most efficient machine learning model is conducted to optimize the prediction accuracy of the selected model. Bayesian hyperparameter optimization using Gaussian process is used to construct the probability model of the objective function and use it to choose the most favourable hyperparameters for evaluation in the real objective function. Since Catboost regressor shows the lowest RMSE and the best prediction accuracy, the best combination of parameters for rate of learning, dense layer numbers, nodes number corresponding to each layer, activation function, and dropout rate are tuned for further optimization of its prediction accuracy. To further verify the validity of the model prediction, a refined numerical modelling using LS-DYNA is employed to simulate the interfacial fracture process. Then the defined bond-slip model is fed as input to the contact relationship among FRP and concrete, and the recovered load-slip curves are used to compare with experimental data to verify the prediction accuracy of the proposed method.