Physics-Parameterized Neural Networks for Non-Destructive Concrete Strength Prediction: A Comparative Study

Prediction of concrete compressive strength using Non-Destructive Testing (NDT) is critical for structural health assessment, but conventional data-driven models often lack physical plausibility and can exhibit degraded performance when extrapolating beyond the training domain. This study introduces and systematically evaluates a hybrid modeling framework belonging to the broader family of Physics-Informed Neural Networks (PINNs), where established concrete maturity models are enforced architecturally. The PINN architecture utilizes a neural network to infer the physical parameters of maturity models from NDT inputs (Electrical Resistivity and Ultrasonic Pulse Velocity), with the final strength prediction governed by the embedded physical equation. A standard Neural Network (standard NN), optimized via a systematic hyperparameter search, and Multiple Linear Regression (MLR) were compared against four PINN variants, each incorporating a different maturity function (Exponential, ACI 209R, FIB Model Code, and Richards). Results from a 5-fold cross-validation procedure demonstrate that while the unconstrained standard NN achieved the highest accuracy in interpolation (R 2 =0.981), its performance was closely trailed by the PINN variants. In the challenging extrapolation task on unseen late-age data, the PINN-FIB model yielded superior performance (R 2 =0.836), demonstrating robust generalization. In contrast, the purely data-driven standard NN failed to generalize (R 2 =-0.071), yielding predictive accuracy inferior to a mean-baseline model, while the MLR model achieved limited generalization (R 2 =0.336). The PINN framework also provides an explainable gray-box mapping from NDT measurements to parameters such as ultimate strength and hydration rate, delivering wider, albeit under-calibrated, predictive uncertainty bounds compared to the standard NN. The findings conclude that enforcing appropriate physical constraints is essential for reliable and transparent predictive models suitable for engineering applications.