Beyond PGA: Role of Tank Geometry and Filling Level in Seismic Fragility of Atmospheric Storage Tanks-a Machine Learning Approach

Abstract

Peak ground acceleration (PGA) is the near-universally adopted intensity measure for seismic fragility assessment of atmospheric storage tanks, underpinning HAZUS, API 650, and Eurocode 8 provisions. Despite its practical convenience, systematic empirical evidence on whether PGA alone captures the full spectrum of damagecontrolling variables has been lacking. This study addresses that gap by assembling and analysing the largest observational damage database reported to date for vertical cylindrical atmospheric tanks, comprising 4614 records from 42 earthquake events spanning 125 years (1900–2024), supplemented with approximately 100 finite element simulation results. A gradient-boosted machine learning framework is applied to derive both PGAonly and multivariate fragility functions for five damage states (DS1–DS5), using stratified five-fold cross-validation and isotonic probability calibration throughout; noting that the reported performance metrics represent upper-bound estimates given the record-level validation strategy. Three principal findings emerge. First, incorporating height-to-diameter ratio (H/D) and filling level alongside PGA raises ROC-AUC by 0.038–0.098 across all damage states, suggesting that geometric and operational variables carry statistically significant information beyond PGA, based on upper-bound cross-validated performance estimates. Second, PGA alone cannot discriminate DS ≥ 4 from DS ≥ 5: Mann-Whitney U and Kolmogorov-Smirnov tests confirm that the PGA distributions of DS4 and DS5 records are statistically indistinguishable (p = 0.631 and p = 0.314, respectively), consistent with the interpretation that structural characteristics — rather than ground motion intensity alone — may govern the transition between extensive damage and collapse in the observational record. Third, the missingness indicator for filling level emerges as the second most influential predictor, suggesting that tanks lacking operational documentation are associated with higher damage probabilities — a statistically robust predictive signal whose physical basis remains to be established, with potential implications for risk-targeted inspection prioritization. Updated lognormal fragility parameters and multivariate ML-MCS curves are provided for all damage states.