Machine Learning is rapidly advancing scientific research and seismology has benefited in this data driven modeling transformation. The current practice of supervised learning approaches is to design task-specific models trained on labeled dataser to achieve high performance on narrow objectives. While effective, these models often lack generalization and robustness, therefore limiting widespread application across different seismic settings and tasks. Seismic waveforms, like audio signals, are continuous time-series data that capture complex patterns of vibrations. This similarity makes seismic data well-suited for approaches originally developed for audio, including foundation model architectures. Foundation models learn a generalized representation from large-scale seismic data self-supervised learning, thus allowing several downstream tasks to be performed in a unified framework. There are several approaches of developing a FM in seismology, from adapting an LLM-style architecture, to leveraging transfer learning from audio AutoEncoder, fine-tuned on seismic waveforms. The learned embedding signals are applied to a set of downstream tasks, such as fault state classification, ground motion regression for early warning, and earthquake phase detection. The results show improvement when comparing to task-specific models, even when applying few-shot learning. The performance of these models highlights both the strengths and current limitations of FMs in seismology, offering valuable insights to guide future efforts in developing more robust and generalizable FMs for this field.

Laura Laurenti is a postdoctoral scholar at ETH-Zurich, where she studies the application of deep learning audio models to seismic data. She received her PhD from La Sapienza University of Rome. Her research includes applying deep learning to laboratory earthquakes, foundation models for seismic data, and diffusion models for seismic data.