Explainable machine learning based seismic response prediction for eccentric bottom frame structures

doi:10.1016/j.istruc.2026.111798

Explainable machine learning based seismic response prediction for eccentric bottom frame structures

2026-04-06

0

PRE

AI

DOI：10.1016/j.istruc.2026.111798

Original

Original request for help

Save

Abstract

En 中文

Bottom frame structures are common in China’s small- and medium-sized cities. These structures typically feature a frame and seismic wall system in the lower stories and a masonry load-bearing system in the upper stories. Irregular stiffness and mass distributions make them susceptible to torsional response and weak-story damage, while limited field observations hinder reliable damage prediction. This study proposes an interpretable framework that combines finite element analysis (FEA) and machine learning (ML). Non-linear time-history analyses were carried out for 113 finite element models with varied inter-story stiffness ratios and bidirectional eccentricities in the first two stories, yielding 339 response samples. Seven inputs were considered: seismic precautionary intensity, inter-story stiffness ratio, PGA, and the X- and Y-direction eccentricities of the first and second stories. Two outputs were predicted, namely the inter-story drift ratios of the bottom frame story and the masonry story. Five ML models were tuned using tree-structured Parzen estimator optimization (TPE) under nested cross-validation. CatBoost performed best for the bottom frame story, while LightGBM achieved the best results for the masonry story. We employed class-conditional CV+ conformal prediction to quantify predictive uncertainty; only the intensity 8 subset did not reach the nominal coverage (90%). Specifically, the X-direction eccentricity strongly correlates with the prediction interval width of the bottom frame story, while PGA shows a similar strong correlation with that of the masonry story. By providing complementary global and local insights, SHAP and LIME revealed attribution patterns that are physically consistent with expected structural response mechanisms.

Keywords:

Bottom frame structures

Seismic response prediction

Machine learning

Inter-story drift ratio

Structural irregularity

Explainable machine learning based seismic response prediction for eccentric bottom frame structures

Abstract

Journal

Organization

Cited Papers