This study introduces a Bidirectional Neural Network (BLiqNet) for predicting 5%-damped spectral acceleration ( S a ) based on seismic source characteristics, propagation path, and site conditions. Unlike conventional Ground Motion Models (GMMs), which rely on deterministic relationships, the BLiqNet leverages a bidirectional learning approach, enabling both forward prediction and inverse inference of key seismic parameters. The model incorporates moment magnitude ( M w ), Joyner-Boore distance ( R J B ), fault mechanism ( F ), hypocentral depth ( H d ), average shear-wave velocity up to 30 m depth ( V s 30 ), and ground motion direction ( d i r ) as input variables, effectively capturing their complex, non-linear dependencies. The model was trained on 23,929 ground motion records from 325 shallow crustal events in the updated NGA-West2 database. The integration of a Liquid Neural Layer and adjoint method enhances interpretability and improves generalization across diverse seismic conditions. Model evaluation through inter-event and intra-event residual analyses confirms its reliability, revealing well-calibrated predictions with minimal bias. The findings highlight the potential of BLiqNet in seismic hazard assessment, site characterization, and ground motion reconstruction, offering a novel data-driven approach for earthquake engineering applications.
Neelamraju et al. (Sun,) studied this question.