Integrated machine learning framework for performance-based seismic assessment of post-tensioned steel-timber hybrid frames with energy-dissipating braces

Performance-based seismic assessment (PBSA) is essential for evaluating both structural safety and post-earthquake recovery. While machine learning (ML) has been used to predict seismic responses, its integration into full PBSA workflows—particularly for loss and downtime estimation—remains limited. This study formalizes a streamlined, building-specific ML-PBSA framework and demonstrates its application to post-tensioned steel–timber hybrid (PTSTH) frames with energy-dissipating braces. Seismic input uncertainty was addressed using Latin hypercube sampling and ground motion selection by return period. Feature engineering and ML algorithms were jointly optimized via tree-based Bayesian method. Surrogate models achieved high predictive accuracy (R² = 0.940–0.966) for key response parameters, with a 95th percentile model improving upper-bound prediction of residual inter-story drift. Fragility analysis confirmed the suitability of ML outputs for damage assessment. The ML models were further integrated into the full PBSA process, including seismic loss and downtime estimation. Results closely matched those from nonlinear time history analysis, with prediction errors under 5 % for seismic events with 975-year return periods or longer.