Evaluation of the seismic resilience index using machine learning based on loss function: a case study of a hospital in Iran

This study evaluates the seismic resilience index for a hospital building in Iran using machine learning (ML) surrogates trained on incremental dynamic analysis (IDA) results. Twenty-two far-field ground motion records were scaled to 20 intensity levels, generating 440 IDA runs. Seven input features were used: magnitude (Mw), focal depth, site condition (NEHRP class III), peak ground acceleration (PGA), spectral acceleration at the fundamental period (Sa(T1)), maximum interstory drift ratio, and peak floor acceleration. The resilience index (Level II seismic hazard) was computed from a functionality‑based recovery function, with Damage State 3 (DS3: interstory drift > 2.5%) mapped to a 45% immediate functionality loss and linear recovery over 90 days. Four ML algorithms – XGBoost, Random Forest (RF), Decision Tree (DT), and Support Vector Machine (SVM) – were trained on 80% of the data and tested on 20% (5‑fold cross‑validation). The predicted resilience indices were 0.502 (XGBoost), 0.540 (RF), 0.519 (DT), and 0.481 (SVM). Random Forest achieved the highest test R² (0.963) and lowest RMSE (0.032). The hybrid IDA‑ML framework provides computationally efficient, interpretable resilience predictions for critical infrastructure.