Physics-guided long short-term memory network for stress–strain characteristic model of fiber-reinforced polymer confined cylindrical concrete

This study presents a predictive model for the complete stress–strain response of fiber-reinforced polymer (FRP)-confined cylindrical concrete using a sequence-to-sequence machine learning framework. Unlike previous studies that focused only on peak strength or strain, this work predicts the entire stress–strain curve, covering both hardening and softening regions. A comprehensive dataset of experiments comprising both synthetic and natural FRPs was compiled for use in the modelling process. A Long Short-Term Memory (LSTM) network integrated with a physics-guided neural network (PGNN) was developed to embed confinement mechanics into the training process. The model achieved an overall R 2 of 0.886, with higher accuracy for synthetic (R 2 = 0.959) than natural FRPs (R 2 = 0.840). Among materials, carbon FRP (R 2 = 0.960) and jute FRP (R 2 = 0.915) performed best. Explainable artificial intelligence analysis identified unconfined compressive strength, ultimate strain, and FRP type as key predictors, while natural fibers such as water hyacinth and cotton showed minor effects. The results demonstrate that the proposed physics-guided framework can reproduce the complete nonlinear stress–strain response of concrete confined with different FRP systems. Compared with the plain LSTM, the physics-guided formulation yields a modest improvement in global predictive accuracy while producing more physically admissible responses. Together with the feature-importance analysis, these findings highlight the potential of physics-guided deep learning to support accurate and interpretable data-driven modeling of nonlinear structural materials.