Cluster Analysis of Long-Period Ground-Motion Simulation Data with Application to Nankai Trough Megathrust Earthquake Scenarios

We developed a clustering method combining principal component analysis and the k-means algorithm, which classifies earthquake scenarios based on the similarity of the spatial distribution of earthquake ground-motion simulation data generated for many earthquake scenarios, and applied it to long-period ground-motion simulation data for Nankai Trough megathrust earthquake scenarios. Values for peak ground velocity and relative velocity response at approximately 80,000 locations in 369 earthquake scenarios were represented by 15 principal components each, and earthquake scenarios were categorized into 30 clusters. In addition, based on clustering results, we determined that extracting relationships between principal components and scenario parameters is possible. Furthermore, by utilizing these relationships, it may be possible to easily estimate the approximate ground-motion distribution from the principal components of arbitrary sets of scenario parameters.