Explainable AI-based framework for 3D structural as-built modeling of aging buildings in Korea using laser scanning

Purpose This study proposes a framework for automatically generating 3D structural as-built models of aging buildings by predicting the structural dimensions of members concealed behind finishes using laser scanning and machine learning, particularly when original construction drawings are unavailable. Design/methodology/approach Laser scanning captures building geometry but includes both structural components and non-structural finishes (e.g. plaster), making it unsuitable for directly extracting structural dimensions. To address this, the proposed framework extracts geometric features from point clouds and employs machine learning techniques to predict structural dimensions of columns, beams, walls and slabs. Six machine learning models – Random Forest, XGBoost, LightGBM, CatBoost, Linear Regression and Support Vector Regression (SVR) – were trained using data collected on 1,560 column elements, 2,875 beam elements, 870 wall elements and 148 slab elements across 50 buildings and validated through two real-world case studies. Findings The best-performing model achieved MAPE of 10.51% for column width (95%CI: 8.47–12.71), 12.98% for column depth (95% CI: 10.86–15.41), 14.91% for beam width (95% CI: 13.63–16.15), 18.91% for beam depth (95% CI: 17.94–19.90), 10.42% for walls (95% CI: 8.89–12.00) and 10.70% for slabs (95% CI: 7.11–14.50). The predicted dimensions enabled reconstruction of 3D structural as-built models without relying on original drawings; the estimates are intended for planning-stage use. Practical implications The framework provides a non-destructive, full-coverage estimate of structural member dimensions that would otherwise require original drawings or local removal of finishes. In practice, it is intended to reduce reliance on destructive investigation and, through the SHAP analysis, indicates where confirmatory openings could be prioritized, thereby supporting planning-stage decisions for the maintenance and retrofitting of buildings whose drawings are missing or outdated. Originality/value This study is among the first to formulate the discrepancy between laser-scanned surface geometry and concealed structural dimensions as a prediction problem and to propose a framework that returns the predicted dimensions to a 3D structural as-built model. Machine learning provides the enabling component, and explainability is used to direct in situ verification rather than as a methodological contribution in itself.