Alternative ground motion model for CENA region using a deep neural network integrated with transfer learning technique

Empirically derived ground motion models (GMMs) that predict ground motion parameters of engineering significance have become unreliable in the data sparse regions due to lack of a comprehensive dataset. As a result, alternative methods have been explored to develop more robust GMMs for these areas. In this context, a non-parametric deep learning approach, associated with a machine learning technique called transfer learning, has been shown to deliver more accurate estimates for data-scarce active regions. In the current study, a deep neural network (DNN) prediction model with transfer learning is derived for a stable-continental region such as Central and Eastern North America (CENA), using data from an active region like Western North America (WNA). Ground motion intensity measures, including peak ground acceleration (PGA), peak ground velocity, and 5% damped pseudo spectral acceleration, are estimated using earthquake parameters such as magnitude (Mw), focal depth (D), focal mechanism (FM), Joyner-Boore distance (Rjb), and shear wave velocity of the top 30 m of soil (Vs30). The WNADNN model is initially trained with the NGA-West2 database, which is then used to initialize the training of the CENADNN with the NGA-East database. The predictions from the model, along with the resulting standard deviations, agree well with regional data and other GMMs derived for the region using various methodologies. Therefore, the transfer learning technique proves effective in deriving prediction models by using datasets from regions with different tectonic settings.