An interpretable deep learning Vertical Ground Motion model based on NGA-West2 database

Compared to traditional empirical models, machine learning-based ground motion models offer superior accuracy and reliability. However, the "black-box" nature of machine learning techniques limits the interpretability of these models, underscoring the need for interpretable machine learning-based approaches. In this study, we developed a vertical ground motion model for peak ground acceleration (PGA), peak ground velocity (PGV), and 5%-damped pseudo-spectral acceleration (PSA) (ranging from 0.01 s to 10 s) using the interpretable Neural Additive Models (NAMs) algorithm. The model was trained on 14,995 vertical ground motion records from 257 seismic events in the NGA-West2 database. We conducted a comprehensive evaluation of the model, including performance analysis, physical characterization, residual analysis, and interpretability analysis. The results demonstrate that the proposed model strikes an effective balance between accuracy and reliability while maintaining the physical characteristics of traditional models. Furthermore, the model offers excellent interpretability: the influence of input features on the model varies, with magnitude and rupture distance being the most significant contributors. Specifically, the impact of magnitude increases with its value, while the effect of rupture distance diminishes as it increases. In contrast, the depth to the top of the rupture and time-averaged shear-wave velocity in the top 30 meters of soil have relatively minor effects, which slightly decrease as their values rise. Moreover, the model enables interpretable analysis of the PSA spectrum for various ground motion input parameters, allowing for the identification of their respective impacts on the spectrum.