Ground motion pulse-like detection and classification: combining different approaches for comprehensive metadata assignment supporting ML techniques for engineering applications

It is well known that near-field earthquake ground motion can be characterized by strong velocity pulses that may cause extensive damage to buildings and structures, as recently documented for the Mw 7.8 and Mw 7.5 earthquake doublet of the 2023 Turkey seismic sequence. 
Usually only directivity pulses are investigated, neglecting other characteristics such as unilateral/bilateral shape, presence of multiple-pulses as well as other features that can support classification of pulse causes. As observed in  recent studies on the directivity pulses of the 2023 Turkey seismic sequence (e.g. Yen et al., 2025), this practice leads to a significant variability in the pulse properties of the observed records, highlighting that factors  beyond rupture directivity also play a crucial role in shaping pulse characteristics, such as  site effects, permanent ground displacements, local heterogeneities in slip amplitude, orientations, and fault kinematics.  
In this study, we provide a methodology that combines different approaches (Baker et al., 2007; Shai and Baker, 2011, 2014; Ertruncay and Costa, 2019; Chen et al., 2023; Chang et al., 2023) for pulse detection and classification. The aim is twofold: on one hand, we aim to extend  metadata assignment for a better characterization of pulse properties; on the other hand, we provide a ML-ready dataset to support development of advanced ML techniques for pulse classification. Indeed training of ML-based algorithms needs the availability of large labelled high-quality dataset. For this purpose, we exploit two comprehensive worldwide datasets of near-source records: the NESS2.0 (Sgobba et al., 2021), which collects real earthquake records, and the BB-SPEEDset (Paolucci et al., 2021), consisting of  ground motion data from 3D Physics-Based Numerical Simulations.