Machine learning-based framework for real-time seismic damage assessment of RC buildings in Nepal

Abstract

Seismic fragility curves help assess structural damage based on a few structural and ground motion (GM) features. In Nepal, diverse construction practices exist, and current standards require reinforced concrete (RC) buildings to ensure structural integrity under design spectra for different soil types. To study this, 1942 low-rise RC building typologies representing various construction practices were modeled in OpenSees and subjected to 28 soil-specific GMs for nonlinear time history analysis (NLTHA) to estimate maximum inter-storey drift ratio (MIDR). The resulting damage distribution across construction practices and soil types was analyzed. Furthermore, five practice-specific and one generalized machine learning (ML) model were developed, integrating structural features (such as member dimensions, geometry, material strengths, and reinforcement details) and seismic features (including GM intensity, duration, frequency content, and energy indicators) to estimate MIDR. The models achieved mean absolute errors between 0.0258 and 0.1506, and coefficients of determination ( R 2 ) between 98.92% and 99.80%, demonstrating robust performance. Feature importance and interdependence were evaluated using Shapley values and SHapley Additive exPlanations (SHAP) dependence plots, identifying GM mean period, predominant period, and natural vibration period as the three most influential and interacting parameters governing building response and damage prediction.