Enhancing seismic demand assessment through explainable artificial intelligence and Predictive Power Score-based ground motion parameter selection

Abstract

Effective seismic risk management requires reliable, interpretable, and data-driven models to navigate the highly stochastic and uncertain characteristics of earthquake ground motions. Conventional ground motion intensity measures (GMIMs) often provide limited decision support in seismic demand modeling because they inadequately represent the complex, nonlinear, and interaction-driven relationships between seismic inputs and structural vulnerabilities. To address these uncertainties, this study presents an explainable artificial intelligence (XAI) framework that employs the Predictive Power Score (PPS) as a robust, data-driven metric to support hybrid ground motion parameter selection for seismic demand assessment. The framework is evaluated using a large-scale dataset comprising 492,096 nonlinear time history simulations generated from 466 ground motion records and 1,056 structural models. Twenty scalar GMIMs and 190 hybrid parameter combinations are assessed based on their ability to predict displacement demand, a key indicator of seismic risk. Results indicate that the Velocity Spectrum Intensity and Effective Design Acceleration hybrid achieves the highest predictive performance with a PPS value of 0.674, representing a 14.4% improvement over the best scalar measure (VSI, PPS = 0.589, p < 0.001) and a 20.7% reduction in mean absolute error. Statistical robustness is confirmed using Wilcoxon signed-rank tests and 10,000 bootstrap resampling iterations. Model transparency is ensured using Shapley Additive Explanations (SHAP), which unravel the "black-box" nature of the predictions and reveal physically consistent, synergistic relationships between ground motion characteristics. Overall, the framework strengthens risk-informed decision-making and stochastic demand assessment for low- to mid-rise reinforced concrete (RC) buildings in shallow-crustal seismic zones, significantly reducing predictive uncertainties.