Machine learning-based prediction of carbonation-induced reinforcement corrosion initiation risk in reinforced concrete under Kinshasa climate conditions

Natural carbonation of reinforced concrete is a major mechanism of steel depassivation and may lead to corrosion initiation once the carbonation front reaches the reinforcement, making it a critical durability issue for structures exposed to tropical urban climates. This study developed a machine learning framework to predict carbonation-induced reinforcement corrosion initiation risk under climate conditions representative of Kinshasa. The workflow combined a RILEM natural carbonation database containing 1,744 records, including 863 mix-level observations and 6,879 time-series measurements, with Kinshasa climate descriptors based on temperature, relative humidity, and an atmospheric CO₂ proxy. After filtering, 847 observations contained usable numerical climate information for model training. Four models were evaluated, namely Ridge regression, Random Forest, Gradient Boosting, and a median-based dummy baseline. Gradient Boosting achieved the best performance, with grouped cross-validation by reference yielding an RMSE of 0.0737 and an R² of 0.6101, while the reference-based holdout test produced an RMSE of 0.0668, an MAE of 0.0510, and an R² of 0.6488. For a 25 mm concrete cover, the median estimated time for the carbonation front to reach the reinforcement level was 18.30 years under a typical dry Kinshasa climate scenario and 16.05 years under an annual mean Kinshasa climate scenario; at 30 years, the proportion of profiles reaching corrosion initiation was 70.72% and 74.26%, respectively. These results show that an open-data, machine learning-based framework can provide a first quantitative estimate of carbonation-driven corrosion initiation risk in Kinshasa, while also highlighting the need for future local validation and improved representation of the most humid exposure conditions.