Neural Nets: applications to two relevant seismological problems

Italy is among the Mediterranean countries with a high seismic hazard, and risk. Over the past century, it experienced several earthquakes that caused human and economic losses. There is not yet a way to predict earthquakes, but through appropriate monitoring techniques, seismic studies and anti-seismic engineering design, we can enhance our level of prevention against this natural event. In this thesis we aim at providing a contribution to two subsets of the site effect assessment branch of research, and specifically to some aspects related to the Horizontal to Vertical Spectral Ratio (HVSR) technique and Ground Motion Modelling (GMM). To this aim, we employ machine learning techniques, here applied in a not-yet-experimented way. In conclusion, in this thesis we explored a number of different neural network-based approaches, such as artificial neural network, convolutional neural network which includes also U-Net, to different classification or prediction problems in seismology. This was important to learn the advantages and disadvantages of the different approaches and constitutes a first basis to build more robust studies in the future and to expand the applications studied so far. In the first case, we employed Artificial Neural Networks and a pre-trained SqueezeNet to separate stratigraphic from non-stratigraphic H/V peaks. This is important to capture the features of H/V curves that are really linked to the soil properties and not to external interference. Implementing neural networks for HVSR analysis could complement traditional methods (e.g. SESAME criteria), simplifying peak recognition and supporting operators in data interpretation. In the second case, we used a U-Net neural net capable of generating region-specific, fully non-ergodic, and data-driven GMM, an approach not previously explored in the Italian context. Afterwards, we compared the outcomes with traditional ground motion prediction equations. Potential future developments of this methodology could offer shaking maps in areas currently lacking seismic data.