Attention-Enhanced Progressive Transfer Learning for Scalable Seismic Vulnerability Assessment of RC Frame Buildings

Urban infrastructure in seismic zones demands efficient and scalable tools for damage prediction. This study introduces an attention-integrated progressive transfer learning (PTL) framework for the seismic vulnerability assessment (SVA) of reinforced concrete (RC) frame buildings. Traditional simulation-based vulnerability models are computationally expensive and dataset-specific, limiting their adaptability. To address this, we leverage a pretrained artificial neural network (ANN) model based on nonlinear static pushover analysis (NSPA) and Monte Carlo simulations for a 4-story RC frame, and extended its applicability to 2-, 8-, and 12-story configurations via PTL. An attention mechanism is incorporated to prioritize critical features, enhancing interpretability and classification accuracy. The model achieves 95.64% accuracy across five damage categories and an R² of 0.98 for regression-based damage index predictions. Comparative evaluation against classical and deep learning models demonstrates superior generalization and computational efficiency. The proposed framework reduced retraining requirements across varying building heights, shows potential adaptability to other structural typologies, and maintains high predictive fidelity, making it a practical AI solution for structural risk evaluation in seismically active regions.