Application of Deep Learning Technique for Site Amplification at Central and Eastern North America

Creative Commons License

İlhan O., Hashash Y. M. A., Stewart J. P., Rathje E. M., Nikolaou S.

The 6th IASPEI / IAEE International Symposium: Effects of Surface Geology on Seismic Motion, Kyoto, Japan, 30 August - 01 September 2021

  • Publication Type: Conference Paper / Unpublished
  • City: Kyoto
  • Country: Japan
  • Ankara Yıldırım Beyazıt University Affiliated: Yes


This paper presents a new set of site amplification models for Central and Eastern North America (CENA) through deep learning via Artificial Neural Network (ANN). The amplification models are for response spectral (RS) and Fourier amplitude spectral (FAS) ordinates. Compared to functions regressed using the same simulation data, the ANNs reduce bias for certain conditions (e.g., shallow site amplification) and aleatory variability of estimations. To the extent that the simulations may represent real site response, these results illustrate inherent limitations of fitting predefined functional forms to simulated data. In this study, the ANNs trained using an enhanced simulation database, are shown to reproduce attributes of site-specific response (e.g. the features of peak amplification, and the period elongation behavior in nonlinear analyses) and to illustrate fundamental difference between RS and FAS amplification due to the influence of motion frequency content on the former. A standard deviation for ANNs is also proposed.