Abstract
Abstract
Accurate prediction of compressive strength in coconut fiber reinforced concrete (CFRC) is essential for sustainable infrastructure design. While data-driven machine learning models offer high predictive accuracy, they often violate fundamental physical laws of cement hydration and composite mechanics, leading to unreliable extrapolations. This study presents a novel Physics-Informed Neural Network (PINN) framework that explicitly encodes concrete material physics including cement hydration kinetics, Abrams’ water-cement ratio law, recycled coarse aggregate (RCA) degradation effects, and fiber reinforcement efficiency, into the learning architecture. Using a comprehensive dataset of 586 experimental records compiled from 30 independent studies, the proposed PINN architecture integrates physics-engineered features (hydration factors, Abrams indices, fiber efficiency parameters, and RCA reduction coefficients) with a deep neural network regularized by physical monotonicity constraints. The PINN achieved an \(\:{R}^{2}\) of 0.9482, RMSE of 3.89 MPa, and MAPE of 8.69%, demonstrating comparable accuracy to gradient boosting (\(\:{R}^{2}=0.9673\)) while ensuring complete compliance with age-monotonicity constraints and significantly reduced violations of water-cement ratio and RCA replacement laws. SHAP-based explainability analysis reveals that the PINN correctly captures the diminishing returns of fiber content beyond optimal dosage (1.5–2.0%), consistent with micromechanical fiber bridging theory. The framework provides a trustworthy, interpretable, and physically consistent tool for CFRC mix design, enabling reliable extrapolation beyond the training domain.