Physics-Informed Machine Learning for Carbonation Depth Prediction in Concrete

The durability of reinforced concrete structures is significantly affected by the carbonation process, which decreases the alkalinity of the pore solution and initiates corrosion of the steel reinforcement. However, the square roots of time equations, which are Fickian diffusion-based, are not able to accurately capture the nonlinear interactions of material properties with environmental factors. To overcome this limitation, this research introduces a novel hybrid model based on the integration of a physics-informed neural network (PINN) with residual regression via CatBoost, a categorical boosting algorithm. Using an expanded dataset of 6000 samples, the first stage of the model, which is based on the physics-informed neural network, is able to learn the underlying physics of the diffusion process by imposing monotonicity constraints. The second stage of the model, which is based on the CatBoost algorithm, is able to learn the residuals of the nonlinear interactions of factors such as the curing time, water–cement ratio, and supplementary cementitious material reactivity, which are not captured by the underlying physics of the diffusion law. Data augmentation via physics-based resampling increased the dataset from 3000 to 6000 samples. Validation of the model using 1200 samples resulted in R2 = 0.871, MAE = 15.362, and RMSE = 24.37. SHAP confirmed that the model was physically consistent with the principles of concrete technology, reversing the counterintuitive linear correlations to accurately capture the protective effect of longer curing times. The suggested framework offers a practical method for enhancing durability evaluation and aiding the maintenance and service-life management of reinforced concrete structures.