Correlation models for next-generation amplitude and cumulative intensity measures using artificial neural networks

Abstract

Various next-generation intensity measures (IMs) have emerged in recent years. For instance, filtered incremental velocity, FIV3, was shown to be efficient in predicting the collapse of structures, or average spectral acceleration, Sa_avg, shown to be well-correlated with a wide range of response, from initiation of structural damage yielding to structural collapse. Different ground motion models (GMMs) are available to predict the probabilistic distribution of these IMs, given a set of seismological parameters. Nonetheless, a few key correlation models are still missing in the literature, which are elemental for a more holistic approach to ground motion record selection. These models are vital when performing seismic hazard analysis and ground motion record selection using state-of-the-art approaches like the conditional spectrum or generalized conditional intensity measure (GCIM) methods. This study uses a single machine-learning-based generalized ground motion model (GGMM) to estimate all the IMs. The model is based on the Next-Generation Attenuation (NGA)-West2 database and was utilized to quantify the correlations between the residuals. Correlations were calculated for intra- and inter-event residuals, but only those for total residuals are presented here, given their utility. Since the calculation of residuals comes from the same GMM and the same database of ground motions, this produces more consistent correlation coefficients between all IMs. To facilitate the usage of these correlation coefficients, predictive models of the empirical data were developed using machine-learning-based techniques, namely artificial neural networks (ANNs). It was found that FIV3 is strongly correlated with Sa(~1 s) and itself across all periods and has a weak negative correlation with duration at short periods and near-zero correlation for longer periods. Also, a stronger negative correlation between Sa and significant duration was found, compared with other prominent existing models. Direct correlation models between Sa and Sa_avg are also proposed.