Direct seismic data inversion for volcanoes using machine learning: a comparison of 2D and 3D cases

Imaging volcanic interiors is of paramount importance for understanding volcano-seismic signals and their underlying sources. However, determining fine scale structure in highly heterogeneous media is a significant challenge using traditional imaging approaches. Furthermore, modelling and inversion tools often employ cumbersome and lengthy procedures, which can be slow to implement, especially during volcanic crises when results are needed swiftly as large data volumes  arrive at Volcano Observatories. Machine-learning (ML) methods, which have experienced rapid growth over the last decade, have strong potential to address this challenge due to their suitability for complementing physics-based numerical simulations and inversion. In particular, we examine the feasibility of imaging small-scale heterogeneities beneath volcanoes, such as propagating individual dykes, directly from seismic data using rapid ML-based imaging.

Here we build on previous work where a large suite (> 5000) of seismic earthquake gathers (i.e. seismic records from individual earthquakes) derived from numerical simulations in highly heterogeneous 2D velocity models, were used to train a Fourier Neural Operator (FNO). Subsequently that FNO was used to invert for complex structure in previously unseen geologically realistic 2D models. As the training procedure is extremely computationally expensive, and is likely prohibitive in 3D, here we ask: “can meaningful information be retrieved from seismic data derived from 3D simulations, based on an FNO that was trained only on 2D seismic data”?  We see the answer to this question as important, as it helps determine the nature of the FNO training required in order to apply this new methodology beyond the numerical domain into the 3D physical world.

We build 3D models that are consistent with the 2D models used for machine learning training. Seismic data are generated from these models, and we evaluate how well a 2D pre-trained algorithm can recover geological structures and velocity characteristics from the 3D data.