GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 2.1 GNGTS 2024 computational domain for tsunami propagation consisted of five levels of nested grids with increasing resolution approaching the Ravenna harbour (640, 320, 80, 20 and 5 m, respectively). The tsunami simulations were conducted with the Tsunami-HySEA software. Implemented in CUDA (Compute Unified Device Architecture) and parallelized for running in multi-GPU architectures (de la Asunción et al., 2013), the code solves the non-linear shallow water. It has passed proper benchmarking, in particular the National Tsunami Hazard and Mitigation Program, USA (Macías et al., 2017, 2020b, 2020a). The method shown in Fig. 1,a diagram has been adopted. In addition to the Catania test case, we applied the source refinement step. The regional model is again represented by the NEAMTHM18 tsunami hazard model (Selva et al. 2016, Basili et al., 2021). The first massive selection of scenarios involves the application of weighted importance sampling with disaggregation. This operation, carried out with a target near Ravenna, provides a total of about 1500 scenarios of the ensemble composing the regional hazard. The close sources (within the first 125 km from the target location) are perturbed 10 times around their NEAMTHM18 values. The far sources are not perturbed. This operation widens the number of scenarios (about 4000) and considerably enriches the variability of the source: this can be a welcome operation in the passage from a regional description to a local one. The obtained scenarios are the input for the first set of tsunami simulations conducted over 4 (6 for the more distant sources) hours of propagation on the first 3 grids of the bathymetrical domain. The water profiles over 7 points of the 80 m resolution grids, taken at an approximated depth of 10 m, are the inputs for the subsequent filtering conducted with a clustering algorithm. The final set of simulations (on 250 cluster representative scenarios) is run over all the available grids, reaching the higher resolution of 5m in the proximity of the target area, the Ravenna harbour. The results obtained with the fine grid simulations are used to produce the hazard maps for maximum inundation height (Fig. 1,c). Alongside the hazard results, the behaviour of the waves inside the Porto Corsini canal was studied, in order to retrieve useful information for sites and structures near the canal, taking advantage of the use of detailed bathymetry (5 m resolution). Conclusion In this study, we developed a local hazard model based on an existing regional model: this approach significantly reduces the computational workload without sacrificing precision or complexity. The key ingredient to reduce computational load and introduce an innovative strategy is represented by the coupling of the probabilistic insights gained from disaggregation with targeted sampling through importance sampling. This combination allows for an effective selection of scenarios that contribute most to the hazard, without the need for arbitrary and subjective cuts. The resulting ensemble is substantially reduced compared to the initial scenario