Annunziata, Daniela (2025) Computational frameworks for operational seismology: synthetic data generation, adaptive classification, and deep learning for source characterization. [Tesi di dottorato]

Documento PDF ANNUNZIATA_DANIELA_38.pdf Visibile a [TBR] Amministratori dell'archivio Download (34MB) | Richiedi una copia

Abstract

This thesis develops Machine Learning (ML) methodologies to address critical challenges in operational seismology, particularly in Earthquake Early Warning (EEW), and to provide frameworks that remain both generalizable and scalable. The contributions of this work unfold along three main directions. First, the thesis introduces a rapid classification framework that remains robust under dynamic sensor network conditions. It combines a fixed grid-based representation with a heterogeneous ML ensemble, achieving high accuracy and stable performance even under simulated network changes. This makes it suitable as a high-speed decision layer for EEW systems. Second, it addresses data scarcity in source characterization through large-scale physics-based synthetic data generation. A parallel High-Performance Computing (HPC) pipeline produces extensive, physically grounded datasets for focal-mechanism analysis. Third, the thesis introduces DeepFoc, a Deep Learning architecture for direct estimation of seismic source parameters. Trained exclusively on synthetic data generated on-the-fly, DeepFoc explicitly targets the long-standing sim-to-real transfer challenge. Validation on real seismic data from the Campi Flegrei caldera shows that DeepFoc provides more stable and accurate focal-mechanism estimates than traditional inversion approaches, particularly for low-magnitude events. Together, these methodologies offer scalable solutions to key operational challenges in seismic monitoring and demonstrate the practical viability of the sim-to-real paradigm in complex geophysical environments. Their design emphasizes transferability, underscoring the broader relevance of the proposed frameworks.

Downloads

Actions (login required)

Modifica documento