Generation of a Response Spectrum from a Fourier Spectrum Using a Recurrent Neural Network: Application to New Zealand

Ground motion prediction equations (GMPEs) are developed using past strong-motion records to predict the effect of future events. Often, the records in the database are incomplete, not covering all possible input scenarios or not recorded at the site of interest for performing site-specific hazards. Such cases are handled by adjusting the GMPEs to suit the required site/region characteristics. Recent studies have shown that scaling the Fourier amplitude spectrum (FAS) rather than the pseudo-spectral acceleration (PSA) is physically justifiable. The present work develops a recurrent neural network to predict the PSA ordinates from the FAS and duration (D5-95) for the New Zealand region. The developed network has no potential underfit or overfit and has a strong correlation coefficient, R > 0.97, with total sigma values in the range of 0.1-0.13 (log10 units). If the predicted FAS and duration are used as inputs, its uncertainty must be included in the final sigma, which lies from 0.25 to 0.3 (log10 units). At low frequency, scaling of FAS and PSA values is identical. In contrast, scaling of higher-frequency FAS values affects the wide range of the PSA values, with a prominent effect initially observed at lower frequencies and later at higher frequencies.