Assessment of ground motion amplitude scaling using interpretable Gaussian process regression: Application to steel moment frames

Abstract The process of ground motion selection and scaling is an integral part hazard‐ risk‐consistent seismic demand analysis structures. Due to the lack records that naturally possess high amplitude intensity, research community generally relies on match a target hazard intensity level. factors used are frequently as 10. criticism received in previous studies, extent its has become matter debate, various constraints have been proposed. primary argument against unrestricted unrealistic nature scaled possible biases caused engineering parameters (EDPs) This study presents framework utilize machine‐learning statistical techniques for assessment nonlinear time‐history (NTHA) utilizes Bayesian non‐parametric Gaussian regressions (GPRs) surrogate models obtain estimates EDPs unscaled motions. GPR developed based large‐scale five steel moment frames (SMFs) using 200 as‐recorded motions ten spectral acceleration levels, () (ranging from 0 g maximum considered earthquake, MCE) 2500 representing 50 scale (), 10 levels each SMF. For building, two types considered: i) peak inter‐story drift ratio (PIDR) ii) floor (PFA). To provide better interpretation models, concept explainable artificial intelligence (i.e., Shapley additive explanation, SHAP) insights into decision‐making with respect . Then, GPR‐based EDP under corresponding different SFs compared comparison conducted Kolmogorov–Smirnov (KS) hypothesis test. Results indicate range allowable s depends factors: level (characterized by ), dynamic properties building. In general, it noticed between 0.5 3.0 PIDRs, 0.6 2.0 PFAs. Finally, adhered discrepancies observed measures amplitude‐, duration‐, energy‐, frequency‐ content sets