Neural Network-Based Subduction Ground Motion Model and Its Application to New Zealand and the Andaman and Nicobar Islands

ABSTRACT

A deep learning model is developed for the Next Generation Attenuation – Subduction database for predicting spectral accelerations and peak amplitude measures. The developed model satisfies the statistical criteria necessary for prediction. Standard deviations lie in 0.2864–0.3809, 0–0.2696, and 0.4514–0.7892, range for inter-event, -region, and intra-events, respectively. Transfer learning is applied to the New Zealand region. Probabilistic seismic hazard analysis is performed for the Andaman-Nicobar region and obtained a peak ground acceleration of 0.6–0.7 g and 0.4–0.5 g at the Andaman and the Nicobar Islands, respectively, for a 2475-year return period.

KEYWORDS:

• GMPE
• NGA-Sub
• deep learning
• transfer learning
• seismic hazard
• Andaman and Nicobar Islands

Data and Resources

The NGA-Sub data is processed by PEER; New Zealand data is collected and processed by GeoNet. Neural network architecture is plotted in latex. The developed NGA-Sub MATLAB code and the transfer learning New Zealand MATLAB code are also submitted.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Supplementary Material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/13632469.2022.2121333.