Hangyeol Lee ; Hyunwoo Hong ; Dohyeon Kim ; Seungwoo Han
https://dx.doi.org/10.6106/KJCEM.2026.27.1.049
Leakage in multifamily housing originates from rooftop waterproofing layers. When the actual service life falls short of the design life, leakage occurs due to maintenance failures within the effective lifespan. This study develops and evaluates a design-stage machine-learning algorithm to guide the selection of a waterproofing system by predicting the effective service life of rooftop waterproofing while distinguishing between controllable design variables and exogenous environmental factors. Candidate variables were screened via stepwise regression, multicollinearity checks (VIF), and interaction analysis, yielding eight predictors: substrate, protective topping material, upstand height, construction method, rainfall, minimum temperature, maximum temperature, and solar irradiation. A Random Forest model was trained for pattern learning, and bias was corrected with XGBoost. We compared performance with multiple linear regression and CatBoost under a 70/15/15 train?validation?test split with early stopping. Across six waterproofing methods, the ensemble achieved R²=0.89?0.92, MAE=1.25?1.48%, and RMSE=1.71?2.00% of the observed effective life, reducing bias relative to the Random Forest baseline without increasing variance. Case studies showed that SHAP attributions aligned with defect causes, supporting the model’s use for preventive-maintenance planning and life-cycle cost analysis. The approach provides a practical tool for preempting leakage through datainformed design and maintenance scheduling. Future work will expand datasets, target other leakage-prone components, conduct additional external validation, and quantify and calibrate predictive uncertainty.