Physics-Informed Machine Learning for Carbonation Depth Prediction in Concrete.

Area of Science:

• Civil Engineering
• Materials Science
• Computational Science

Background:

• Carbonation significantly impacts reinforced concrete durability by reducing alkalinity and initiating steel corrosion.
• Existing Fickian diffusion models struggle to capture nonlinear interactions between material properties and environmental factors.

Purpose of the Study:

• To develop a novel hybrid model integrating physics-informed neural networks (PINNs) with CatBoost for enhanced durability prediction.
• To overcome the limitations of traditional square root of time equations in modeling complex concrete carbonation dynamics.

Main Methods:

• A hybrid model combining PINNs for physics learning and CatBoost for residual regression was developed.
• Physics-informed neural networks imposed monotonicity constraints to learn diffusion physics.
• CatBoost captured nonlinear interactions of factors like curing time and water-cement ratio.
• Data augmentation via physics-based resampling expanded the dataset to 6000 samples.

Main Results:

• The hybrid model achieved R² = 0.871, MAE = 15.362, and RMSE = 24.37 on a validation set of 1200 samples.
• SHAP analysis confirmed the model's physical consistency and ability to accurately represent the protective effect of longer curing times.
• The model successfully reversed counterintuitive linear correlations, providing a more accurate durability assessment.

Conclusions:

• The proposed hybrid framework offers a practical and accurate method for evaluating reinforced concrete durability.
• This approach enhances maintenance strategies and aids in the service-life management of concrete structures.
• The model provides a significant advancement over traditional methods for predicting carbonation-induced degradation.