Physics-Informed and Data-Driven Prediction of Carbonation Depth and Service Life Reliability of Concrete Bridges

Predicting concrete carbonation depth is essential for durability design and life-cycle management of concrete bridges. This study combines a physics-informed neural network with a calibrated ensemble of classical learners to estimate the carbonation coefficient and map service life reliability. The learning target is derived from the square-root-of-time law for carbonation front propagation. The PINN is trained on the same descriptors as the ensemble and includes a residual penalty from a reaction–diffusion surrogate. Reliability is obtained by propagating the predictive distribution to the time of steel depassivation. For 8204 tests, the calibrated ensemble gives cross-validated errors of 4.44 mm in MAE and 6.69 mm in RMSE (R2 = 0.77) using 60 physically motivated and spline-expanded features with reference encoding for inter-laboratory domain shift. The PINN gives a validation RMSE of 2.79 mm and R2 = 0.86 on the held-out split. The resulting reliability maps separate cover depth and exposure scenarios while retaining uncertainty consistent with residual dispersion.