NEURO-TOMO · Learning the Earth with Neural Operators

Understanding the fine structure of the Earth is critical for mitigating earthquake risk, assessing volcanic activity, and ensuring the stability of sedimentary basins beneath major cities. Current seismic imaging methods can map subsurface structures, but they face major computational challenges with large datasets. Calculations are slow and often miss small-scale features that are crucial for hazard assessment, limiting our ability to exploit the growing volume of seismic data worldwide.Recent advances in artificial intelligence offer a pathway to overcome these barriers. Machine learning can capture the relationship between Earth structure and seismic wave propagation, enabling accurate waves simulations orders of magnitude faster than traditional methods. This makes it possible to explore thousands of Earth models and quantify uncertainty in imaging results, providing confidence levels that were previously too costly to compute.In this project, I will advance AI-based techniques for seismic imaging and apply them to dense datasets at multiple scales. A key target is the MACIV experiment, the largest volcanological deployment to date, with 150 broadband and 600 short-period stations, offering a unique chance to probe intraplate volcanic systems in Europe. I will also test the framework on other networks, including USArray and AlpArray to apply it across diverse geological context.I will then extend these advances to Distributed Acoustic Sensing (DAS), which converts existing fibre-optic cables into seismic arrays. This step requires further development to handle the variable quality and geometry of DAS data. By adapting the workflow to both dense arrays and DAS, the project will deliver a robust framework for seismic imaging that is transferable across environments from volcanoes to urban basins at different scales, demonstrating how cutting-edge AI can unlock new value from the massive seismic datasets already being collected worldwide.