A physics-guided LSTM–PINN approach for aging-dependent strength and durability prediction of PET–RHA concrete

This study proposes a Physics-Guided Long Short-Term Memory–Physics-Informed Neural Network (PG-LSTM–PINN) framework to predict the time-dependent mechanical and durability properties of concrete containing post-consumer Polyethylene Terephthalate (PET) and Rice Husk Ash (RHA). The aim is to combine hydration-based physical knowledge with deep sequential learning to capture the evolution of concrete properties over the validated curing period of 1–90 days. The model was developed using an experimental dataset consisting of ten concrete mixes (M0–M9), tested at curing ages of 1, 7, 28, 56, and 90 days. Six key performance indicators were predicted simultaneously: compressive strength, splitting tensile strength, flexural strength, water absorption, modulus of elasticity, and rapid chloride permeability. The performance of the proposed PG-LSTM–PINN model was compared with three baseline models: LSTM, LSTM–PINN, and PG-LSTM. The proposed model achieved the best predictive performance, with coefficients of determination (R 2 ) above 0.95 for all predicted properties and an average mean absolute percentage error of about 2.8%. Incorporating hydration kinetics as a physics-informed constraint improved the stability of predictions and ensured physically consistent aging trends. Sensitivity analysis also indicated that cement content, PET content, and RHA content are the most influential variables affecting the predicted properties. These results demonstrate that the PG-LSTM–PINN framework can model the aging behavior of PET–RHA concrete and provide accurate predictions of both mechanical and durability properties using a limited experimental dataset The PG-LSTM–PINN framework integrates temporal deep learning with hydration-based physics constraints to model the time-dependent evolution of concrete properties.