GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 2.1 GNGTS 2024 proxy variables, and considers both the maximum wave height of the tsunami and the coseismic displacement at each target point. To cluster the N SR original scenarios in proxy-space into N CL clusters, we assign each scenario to the cluster with the closest centroid based on the selected metric. To assess the convergence of results with an increasing number of clusters, we rely on the Within-Clusters Sum of Squared (WCSS) or inertia provided by the sklearn tool. This enables a quick visualisation of clustering behaviour and facilitates the identification of a suitable minimum number of clusters using the elbow rule. This approach helps determine the minimum clusters necessary for sufficient convergence of results. While larger numbers of clusters may enhance resolution, they come at the expense of increased computational effort. Following the K-means clustering, a total of N CL clusters are identified. These N CL scenarios serve as a starting point for high-resolution simulations involving flooding at the target site. For each cluster, we designate a representative scenario as the one whose water height profile deviates the least from the cluster's average. Assuming similar effects on near offshore points, scenarios within the same cluster can be reasonably regarded as inducing comparable inundation. The associated weight for each scenario representative is determined by summing the weights of all scenarios within that cluster. Without refinement, this sum corresponds to the number of scenarios in the cluster. It's worth noting that while K-means clustering is based on water height profiles, scenarios in the same cluster are expected to exhibit a relatively similar profile shape. However, this doesn't guarantee that all scenarios and clusters are equally representative of the final hazard: and that’s why the sum of all within-the-cluster scenarios rates is considered. Local hazard quantification High-resolution simulations can be performed for all representative scenarios identified by N CL . The outcomes of these simulations will be utilised to generate hazard curves and maps, offering a representation of our local hazard model. The number of simulations N CL is significantly reduced compared to the number of scenarios making up the initial ensemble: this allows resolution to be carried out even in the absence of a significant computing infrastructure. The test case To test the efficiency and validity of the method described, we considered the area of ​ Catania, Sicily, South of Italy. In this test site, numerous numerical simulations have been produced with significant high-performance computing (HPC) resources for recent studies (Gibbons et al., 2020; Tonini et al., 2021). The 32,363 tsunami simulations were conducted through Tsunami-HySEA software over 4 levels of nested grids: one global 0-grid for the open ocean propagation covering the Mediterranean Sea and three local grids (with resolution 160 m, 40 m and 10 m respectively).