Artificial neural network-based ground motion model for next-generation seismic intensity measures.

This paper presents an application of artificial neural networks (ANN) to ground motion modelling. We focused on developing a generalised ground motion model (GGMM) incorporating several seismic intensity measure (IM) types and their inter-IM correlation. These range from classical IMs, such as peak ground acceleration/velocity/displacement, spectral acceleration, and significant duration, to more advanced IMs recently shown to be better descriptors of structural performa nce, such as average spectral acceleration and filtered incremental velocity. Additionally, three different horizontal component definitions were included for the spectral acceleration-based IMs. A total of nine input ground motion causal parameters are required to use the GGMM developed, based on ground motion records from the NGA-West2 database. ANN was used to perform the regression, which differs from the approaches used in many existing ground motion models (GMMs), and gives the possibility to regress all IMs simultaneously in one model. A mixed-effects regression approach was adopted for the regression and the quantification of the inter- and intra-event variability of the GGMM estimation. The correlations between the IMs were also quantified and briefly presented here, which allows for a more refined prediction of seismic shaking and a unified treatment of prediction and IM correlations. This will allow more advanced record selection for non-linear dynamic analyses to be performed, which can consider several facets of ground shaking currently overlooked in many works. We evaluated the performance of the developed GGMM using several metrics and compared it to various existing GMMs developed with either the classical approach or machine learning methods. The results show that the proposed GGMM exhibits very good predictions, especially considering the wide range of IMs tackled. Lastly, this methodology has the flexibility of being able to add more IMs or horizontal component definitions seamlessly. This manuscript presents a novel generalised ground motion model for traditional and next-generation intensity measures (IMs) used in seismic hazard and risk. Some notable points: • Utilization of artificial neural networks to capture the complex interdependencies among multiple IMs. • The flexibility and accuracy of the model allows for more refined predictions for the next-generation IMs. • All the included IMs are simultaneously regressed using a mixed-effects regression procedure. • The proposed model can be used by the user to provide outputs of several IMs, which accommodates ease of use. • Correlations between the IMs were also quantified to offer a more consistent and unified treatment of IM correlations. • Comparisons of the developed model with existing models from the literature showed good performance and reduced dispersion. [ABSTRACT FROM AUTHOR]