GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 2.2 GNGTS 2024 The target points have been randomly selected with a radial uniform distribuCon around the epicentre. To avoid any bias introduced by the training data, the "computaConal" grid is randomly shiped with respect to the epicentre posiCon for each event. The loss funcCon used to train the model is a linear combinaCon of negaCve log-likelihood (NLL) and Frechet incepCon distance ( ): . EffecCvely, the adopted loss can be seen as a Wasserstein distance with a negaCve log- likelihood penalty term introduced to regularise the results and provide a beber connecCon between the mean and standard deviaCon. Results and Conclusions The proposed hybrid method implements a mulC-step approach in which the neural network performs a very specific task: while it sCll maintains some aspects of a black box-like algorithm, the results of this implementaCon are much more interpretable, specifically with the possibility to address the role of the different components in the final result. The use of data augmentaCon is beneficial even in cases where a good amount of recorded data is available to train the models, because the greater control over syntheCc data could allow the development of more balanced datasets that can, in turn, promote the model to learn more useful low-level features while the fine-tuning phase using real data seems promising in training models able to generate more realisCc results. The proposed method proved to be robust to network geometry changes (both in terms of the number of staCons and their spaCal distribuCon) and to noise. The 30 October 2016 6.5 Norcia earthquake has been chosen to benchmark the method against ShakeMap®. Even though it doesn’t represent an exhausCve analysis, the Norcia event, which required mobilisaCon of emergency response, is indicaCve of the behaviour of the method for the archetype of the seismic event it has been developed for, being a strong event recorded by a high number of staCons with good coverage. In Fig. 3, the PGA median and standard deviaCon obtained with the proposed method and ShakeMap® are compared. In the epicentral area, thanks also to the high density of staCons, both methods provide similar results in terms of median values (panels a) and c) in Fig. 3). Considering the standard deviaCons, the hybrid method generates values that are overall more similar, although consistently greater, than ShakeMap® (panels b) and d) in Fig. 3). Given the PGA probability distribuCons at each target point from the hybrid method and ShakeMap® , the map of the overlapping coefficient has been computed (panel e) in Fig. 3) showing great compaCbility between the two methods and thus the quality of the hybrid method. Despite being non-predicCve (i.e. it reconstructs the ground-shaking field to be consistent with the values recorded unCl that moment rather than foresee future ones), the hybrid L FID L = w NLL NLL + w FID FID M w f H f SM OVL = ∫ R min ( f 1 ( x ), f 2 ( x )) d x