Among the variables that govern how well a retrofitted concrete member performs in service, few are as consequential as the bond between the substrate and an externally attached strengthening layer. For CFRP (carbon fiber-reinforced polymer) systems in particular, the ultimate failure load recorded in single-lap or double-lap shear tests has long served as the principal measure of this bond capacity, yet reliable predictions still rely almost exclusively on labor-intensive laboratory campaigns or analytical formulas calibrated solely on normal-weight concrete (NWC). In this study, three predictive modelling approaches namely artificial neural network (ANN), multiple linear regression (MLR), and a genetic algorithm (GA)-calibrated power-law empirical formula, were developed and rigorously evaluated for predicting the ultimate failure load (Pᵤ) of CFRP-to-concrete specimens encompassing both NWC and lightweight concrete (LWC) substrates. A compiled experimental database of 234 specimens was assembled from two independent sources, incorporating six input parameters: bond length (Lₑ), CFRP width (bₑ), CFRP laminate thickness (tₑ), fiber orientation angle (θ), concrete compressive strength (f′ᶜ), and concrete type. The ANN model with a 100–50 two-hidden-layer architecture achieved the highest predictive accuracy on the independent test set, with R² = 0.83, RMSE = 4.47 kN, and MAE = 3.38 kN, and a 5-fold cross-validation R² = 0.87 ± 0.06. The GA calibration produced an explicit power-law empirical formula with test R² = 0.68, while the MLR model achieved R² = 0.55, confirming the insufficiency of linear approaches for this non-linear problem. Doing a one at a time sensitivity analysis with the ANN model, it was possible to pin down the concrete compressive strength as a dominant factor (sensitivity range = 30.1 kN), then the CFRP fiber orientation angle (22.8 kN) and the CFRP width (22.3 kN). These three were basically the most influential parameters on bond strength. For LWC substrates, the predicted bond capacity came out about 12% lower compared to matching NWC specimens. Overall, the results hand practicing structural engineers validated, data driven predictive tools for design and condition evaluation of CFRP retrofitted LWC structures, and they also help close a gap that was still present in the existing literature.
