Phd in DATA SCIENCE

Thesis title: Application of AI to Seismology and Earthquake Physics: Investigating Propagation Path Variations and Denoising, and Paving the Way for Foundation Models

Inadequate data and predictive techniques have historically limited efforts to forecast earthquakes. Recent advances indicate that lab-engineered earthquakes can be predicted using machine learning. We use fault zone acoustic emissions to predict labquakes with deep learning methods, then we introduce an autoregressive forecasting method to predict fault zone shear stress, also expanding the range of lab fault zones studied. By integrating lab results with field observations, we aim to identify earthquake precursors and develop predictive models for tectonic faulting. One study uses waves from the 2016 M6.5 Norcia seismic sequence, employing DL to differentiate foreshocks, aftershocks, and time-to-failure. A 7-layer CNN achieves over 90% accuracy in classifying seismograms, underscoring DL's ability to track fault zone properties and evolution. Seismic waveforms, rich with information about the earthquake source and geological structures, require effective denoising techniques. We develop a novel CDiffSD: Cold Diffusion Model for Seismic Denoising, outperforming traditional methods by addressing non-Gaussian seismic noise. This model significantly advances seismic data denoising, enhancing waveform analysis accuracy. Further advancements involve SeismicAE: a Seismic Waveform Auto-Encoder by transferring and finetuning learning from the audio autoencoder. This model excels in trace reconstruction, fault state classification, and ground motion regression, significantly improving few-shot training conditions. SeismicAE is a starting point for developing a foundation model for seismology. These advancements in ML and DL establish new standards in seismic data analysis, advancing earthquake forecasting and hazard mitigation strategies.