Artificial neural network-based non-parametric ground motion models for multiple intensity measures in Türkiye

This study proposes artificial intelligence-based nonparametric ground motion models (GMMs), using a single, monolithic artificial neural networks (ANN) with multi-output architecture, to predict intensity measures for Türkiye. Existing regional GMMs are often constrained by limited datasets, restrictive functional forms, and amplitude-based measures, motivating development of models with expanded predictive capability. GMMs are developed using an updated dataset comprising more than 12,000 records from earthquakes between 1976 and 2023. Seismological parameters include moment magnitude ( M w ), Joyner-Boore distance ( R JB ), averaged shear wave velocity in the top 30 m ( V S 30 ), focal depth ( FD ) and fault mechanism ( FM ). The framework is coupled with a mixed-effects regression scheme that explicitly separates total variability into inter-event and intra-event components, enabling a robust representation of source- and site-related uncertainties. The single multi-output model predicts peak ground acceleration ( PGA ), peak ground velocity ( PGV ), Arias intensity ( I a ), cumulative absolute velocity ( CAV ), significant durations ( D 5%–75% and D 5%–95% ) , uniform duration ( T m ), and 5% damped pseudo-acceleration response spectra ( PSa ) over periods 0.03-4.00 s. GMMs are applicable to shallow earthquakes with M w between 4.0 and 7.8, R JB up to 200 km, and V S 30 less than 1500 m/s. Comparisons against empirical models demonstrate improved performance, with lower root mean square error ( RMSE ), scatter index ( SI ), reference index ( RI ), and higher coefficient of determination ( R 2 ), correlation coefficient ( r ) and a 20- index . Sensitivity analysis identifies M w , V S 30 , and R JB as the most influential inputs, confirming ability of multi-output ANN to model complex behaviours for hazard assessment. A graphical user interface is provided for practical use.