PSI - Issue 70

Rachit Sharma et al. / Procedia Structural Integrity 70 (2025) 386–393 Sharma and Laskar/ Structural Integrity Procedia 00 (2025) 000 – 000

388

3

bias of 0.0033. However, the model relies on a basic algorithm and can be enhanced by incorporating ensemble models. These models can be interpreted using the unified SHapley Additive exPlanation (SHAP) approach, which ranks input features owing to the individual influence on the predicting variable. 2. Research Significance Present study explored efficacy of machine learning model namely SVR and ensemble machine learning model XGBoost to estimate FRP-RC beams shear strength. To achieve this, a comprehensive database consisting of 402 experimental data points has been assembled. The ML model parameters have been trained using 10 K -fold cross optimization using GridSearch algorithm. The proposed ML models (SVR and XGBoost) are compared to empirical methods available in design guidelines. Subsequently, the best performing ML model is analysed using SHAP approach to rank the major parameters affecting shear capacity prediction. 3. Experimental Database A dataset comprising 402 FRP-RC beams without transverse reinforcement has been compiled from prior experimental studies, see Appendix A. A total of seven parameters namely web width (b) , effective depth of section (h) , shear span to effective depth ratio (a/d) , clear span (L) , concrete compressive strength (f c ) , longitudinal reinforcement ratio ( ) and FRP elastic modulus (E f ) have been considered based on the existing code provisions and literature (Ali et al. (2021); El Zareef et al. (2021). Figure 1 highlights considered input parameters in which the first row highlights the histograms of input parameters, and second row shows the scatterplot of each parameter with respect to the experimental shear capacity (V exp ) . The shear force due to self-weight has been considered in calculating the shear strength. The cube compressive strength of concrete has been converted to cylinder compressive strength by using a factor 0.8 (European Standard (2023)). Furthermore, in the absence of concrete elastic modulus, the elastic modulus has obtained as 4700 √f c (ACI Committee 318 (2019)). A statistical distribution of the input variables is presented in Table 2 along with the induvial range of input variable. This database convers the FRP-RC beams with various ( a/d ) ratios from 0.5 to 9.60 including both deep as well as slender beams.

Fig. 1. Details of Database of FRP-RC Members Used in Present Study.