Adaptation of the NGA-WEST2 Ground Motion Prediction Equations by Implementing the Resampling Analysis Algorithm

The realistic estimation of Peak Ground Acceleration (PGA) is crucial for the purpose of seismic design in high seismic prone regions. The common practice is using Seismic Hazard Analysis (SHA) to estimate the design spectra in order to be used in the design and rehabilitation of structures. The most influential part of any SHA is the use of Ground Motion Prediction Equation (GMPE), which usually has the highest level of epistemic uncertainty. The Pacific Earthquake Research Centre (PEER) have released two sets of GMPEs, which are known as NGA-WEST1 and NGA-WEST2, and they are introduced as global GMPEs for all regions around the globe. However, the reliability of GMPEs needs to be assessed properly. Therefore, a recent methodology by Azarbakht et al. (2014) has been implemented in this study in order to enhance the given GMPEs only in the case of PGA. The better model in this approach is the one which has less sensitivity due to the small changes in the input catalogue. This effect cannot be captured by the common statistical tests that are widely using in the development of GMPEs. Therefore, three NGA GMPEs are taken into consideration, and the coefficients are optimized by aiming at maximizing the reliability, i.e. Campbell and Bozorgnia, Abrahamson and Silva, and Boore and Atkinson GMPEs. The ground motion database of the Campbell and Bozorgnia (2014) was used throughout this study, which consists of 15493 records of 319 earthquake events. A multi-objective Genetic algorithm was used to optimize four fitness functions, three of them related to resampling of the data and the forth is taken as the LLH. The results show that the employed resampling analysis show better performance when compared to other statistical approaches such as Var-test, Lillifors, and Z-test. However, the optimized coefficients show better GMPE performance with those statistical tests. Error estimation approaches were also considered, i.e. RMSE, MAE, R-square and E methods. In the end, the hazard curves for a hypothetical site are calculated based on the original and optimized GMPEs. The comparison between the obtained hazard curves shows that the hazard curves obtained from the optimized coefficients result in conservative when compared to the hazard curves from the original GMPEs. In conclusion, the optimized GMPEs show better performance when compared to the original GMPEs by means of the common statistical approaches as well as the new resampling algorithm. This proves that the sensitivity of GMPEs to the input catalogue is a key criterion when developing a new GMPE, otherwise the estimated parameters such as PGA will not be accurate enough.