Steel corrosion is a critical deterioration phenomenon in concrete structures, particularly under harsh environmental conditions involving carbonation and chloride ingress. Accurate corrosion rate prediction is essential for structural life assessment and maintenance planning. The existing methods include empirical and electrochemical models; they suffer from a limited adaptability or a high computational complexity. Machine learning (ML) approaches show promise but typically focus on point estimates, while neglecting prediction uncertainty, which is crucial for reliable engineering decision-making. This study was to develop a Gaussian Process Regression (GPR) framework that could integrate physical prior knowledge with uncertainty quantification for steel corrosion rate prediction in cementitious materials, addressing two key questions, i.e., 1) how to enhance GPR accuracy and robustness for corrosion prediction, and 2) whether GPR uncertainty outputs could effectively support corrosion risk classification and engineering decisions.

The study utilized 180 experimental datasets from mortar samples embedded steel exposed to 6 relative humidity (RH) conditions after carbonation treatment. Each dataset contained 15 corrosion-related features categorized as mixture parameters (i.e., cement, slag, fly ash, and silica fume proportions, water-to-binder ratio, and chloride content), material properties (i.e., free chloride content, pH value, and porosity), environmental parameters (i.e., RH, saturation, and water content), and electrochemical parameters (i.e., corrosion potential, electrical resistivity, and [Cl^-]/[OH^-]).

The GPR model incorporated physical prior knowledge via setting the mean function as a linear relationship with a logarithmic chloride-to-hydroxide ratio, i.e., $m(x)=a\cdot \lg \left(\left[{\mathrm{C}\mathrm{l}}^{-}\right]/\left[{\mathrm{O}\mathrm{H}}^{-}\right]\right)+b$, based on the electrochemical corrosion theory. A Gaussian radial basis function served as a covariance function to capture smooth corrosion behavior variations.

The performance evaluation using 10-fold cross-validation demonstrates that the physical prior GPR consistently outperforms conventional GPR across all the feature combinations. The model shows a superior prediction accuracy (i.e., higher R²), a better uncertainty quantification (i.e., lower negative log-likelihood), and more reliable confidence intervals (i.e., higher prediction interval coverage probability).

The corrosion rates are classified into four levels, i.e., low (<0.1 μA/cm²), low-medium (0.1–0.5 μA/cm²), medium-high (0.5–1.0 μA/cm²), and high (>1.0 μA/cm²). The model achieves a high accuracy in identifying severe corrosion conditions, with most of the 159 high corrosion samples correctly classified, though some confusion exists between intermediate levels.

An uncertainty-driven decision strategy is developed using confidence interval width (threshold: 4 μA/cm²) combined with predicted corrosion levels to generate four response categories, i.e., immediate intervention (i.e., high corrosion + low uncertainty), manual review (i.e., high corrosion + high uncertainty), continuous monitoring (i.e., low corrosion + high uncertainty), and accept current condition (i.e., low corrosion + low uncertainty). Approximately 36% of samples are flagged for review or monitoring due to the high uncertainty, demonstrating a value of uncertainty assessment in risk management.

This work presented a framework combining physical knowledge with probabilistic modeling for trustworthy AI-based corrosion prediction. Key contributions could include 1) effective integration of electrochemical mechanisms into GPR architecture, improving model robustness and interpretability; 2) development of a joint corrosion level and uncertainty classification system; 3) establishment of an uncertainty-based engineering decision framework; and (4) validation that uncertainty quantification effectively could identify potentially problematic predictions requiring additional review.

The framework could advance AI applications in structural health monitoring from simple prediction tools to comprehensive decision support systems, providing a foundation for reliable, interpretable infrastructure maintenance strategies. The approach demonstrated a promising potential for enhancing structural safety through improved corrosion risk assessment and management in challenging service environments.