Seismic damage prediction and ground motion characterization in RC buildings via unsupervised learning

This study investigates the latent relationship between ground motion parameters and seismic damage in reinforced concrete (RC) buildings using an unsupervised machine learning framework. A dataset comprising 21 seismic parameters from 466 ground motion records and the nonlinear response of 1,056 RC building models was analyzed. Unlike traditional regression-based studies, this research employs a "blind" learning approach to determine if clustering algorithms can autonomously identify physical damage mechanisms. The K-means algorithm, validated by K-fold cross-validation and internal indices, identified three distinct hazard categories. Crucially, the algorithm autonomously isolated a "Pulse-Like/Resonance-Critical" cluster driven by frequency-content parameters (Vmax/Amax and Tm) rather than simple peak intensity. Furthermore, Principal Component Analysis (PCA) revealed that a reduced "Core Parameter Set" of six indices (including EDA, Arias Intensity, and Vmax/Amax explains 89% of the variance, offering a practical subset for ground motion selection. Finally, manifold learning techniques (t-SNE and UMAP) demonstrated that the complex seismic damage landscape can be unfolded into a quasi-linear manifold, validating the stability of the proposed clustering.