Prediction of the Ultimate Impact Response of Concrete Strengthened with Polyurethane Grout as the Repair Material

The monolithic composite action of structures relies on the interface bond strength between concrete and the repair material. This study uses explainable deep learning techniques to evaluate the ultimate strength capacity (Us) of U-shaped normal concrete (NC) strengthened with polyurethane grouting (PUG) materials. Machine learning algorithms (ML) such as Long Short-Term Memory (LSTM), Random Forest (RF), and Wide Neural Network (WNN) models were developed to estimate Us by considering five input parameters: the initial crack strength (Cs), thickness of the grouting materials (T), mid-span deflection (λ), and peak applied load (P). The results indicated that LSTM models, particularly LSTM-M2 and LSTM-M3, demonstrated superior predictive accuracy and consistency in both the calibration and verification phases, as evidenced by high Pearson’s correlation coefficients (PCC = 0.9156 for LSTM-M2) and Willmott indices (WI = 0.7713 for LSTM-M2), and low error metrics (MSE = 0.0017, RMSE = 0.0418). The SHAP (SHapley Additive exPlanations) analysis showed that the thickness of the grouting materials and maximum load were the most significant parameters affecting the ultimate capacity of the composite U-shaped specimen. The RF model showed moderate improvements, with RF-M3 performing better than RF-M1 and RF-M2. The WNN models displayed varied performance, with WNN-M2 performing poorly due to significant scatter and deviation. The findings highlight the potential of LSTM models for the accurate and reliable prediction of the ultimate strength of composite U-shaped specimens.