A NEW COMPUTATIONAL METHOD FOR SEISMIC HAZARD ANALYSIS Grace S. Wang, Chaoyang University of Technology, Taiwan F. K. Huang, Tamkang University, Taiwan ABSTRACT The purpose of this paper is focused on the determination of the seismic hazard for Taichung Area in Taiwan, based on the recorded historical data provided by National Central Weather Bureau of Taiwan. Before performing the probabilistic seismic hazard analysis (PHSA), the parameters related to the seismic input are identified based on the above data. In the conventional seismic hazard analysis, it is necessary to predetermine the distributions of the associated parameters involved in the probabilistic seismic hazard analysis. In order to overcome this shortcoming, the paper proposes a new methodology of determining the distributions by the concept of Kernel sampling density estimation. This approach is more nonparametric in the sense that less rigid assumptions will be made about the distribution of the observed data. The data is used to determine the Kernel sampling density function of the parameters and then the Monte Carlo simulation method is employed to estimate the associated seismic hazard. However, for the case of low seismic hazard, the required computational cost to reach an accurate result may be expensive. Therefore, simple Kernel method of importance sampling methods is performed with an aim to reduce the statistical error inherent in Monte Carlo methods. 1. INTRODUCTION The inadequate behaviour, during recent seismic events, of buildings designed according to current earthquake-resistant design codes, has given place to intense discussion regarding the needs to revise these codes, and the way in which earthquake- resistant design is currently conceived (i.e., current methodologies). One concept that can help in formulating and transparent earthquake-resistant design approach is that of performance-based design. In performance-based design, the desired behaviour of the building during ground motions of different intensities (design objectives), should be established in a qualified manner, before the numerical phase of the overall design process starts. In this regard, we have to determine levels of ground shaking, usually defined in PGA, with respect to different return periods or annual exceeding probabilities. Probabilistic seismic hazard analysis can provide a curve that represents the annual exceeding probability with respect to different intensity level of ground motion. In this sense, seismic hazard analysis plays an important role. The exceeding probability can be expressed in a way similar to system reliability as ∫ = f D F d f P x x X ) ( (1) in which X the random vector composed of the parameters involved in the seismic hazard, the associated joint density function, the domain of integration where the specified level of peak ground acceleration (PGA) is exceeded. If the number of dimensions is small, the numerical integration of equation (1) may be possible. However, in general, is a multi-dimensional density function with complex mathematical form and the integrated domain is irregular with highly nonlinear boundaries. Hence an analytical calculation of the exceeding probability defined in equation (1) is usually impossible or impractical. This leads to the need for approximate approaches. Simulation methods, particularly Monte Carlo simulation methods, are widely employed not only in mathematics, but also in the problems of system reliability and seismic hazard. Their usefulness is commonly recognized in solving not only integral equations but also differential equations. Indeed, the use of Monte Carlo simulation to estimate is a practical issue. On the other hand, it is necessary to treat the model parameters, such as the focal depth, the earthquake magnitude, and the maximum magnitude, as the random variables. In the conventional probabilistic seismic hazard analysis, more effort is devoted to predetermine the distribution of the associated variables. This can be done by first specifying some distribution to the variable, followed by estimating the parameters involved in the distribution using the ) (x X f f D F P ) ( x X f