Predicting near-surface seismic data and velocity models using synthetically-trained deep learning methods: applications in data-rich environments

Recent advances in machine learning offer a new way for Earth Scientists to make predictions about geological subsurface properties. In particular, methods based on Fourier Neural Operators (FNOs) are increasingly being used as a substitute for conventional approaches based on numerical forward modelling and inversion, at a fraction of the computational cost. Most importantly, FNOs have been shown to predict accurate 2D and 3D forward modelling simulations of seismic waves up to several hundred times faster than physics-based solvers after training.

In synthetic volcanic settings to date, FNOs have been applied successfully to both the forward and inverse problem, capturing the fine-scale velocity structure of heterogeneous models. However, transferring the successful performance of simulation-trained FNOs to make accurate predictions from field-gathered seismic data is yet to be achieved. In order to accomplish this for volcanological data, training models would need to contain representative small-scale velocity heterogeneities and topography in order to produce highly scattered codas in the synthetic seismograms.

This research presents work in progress on simulation-to-real applications of FNOs using field-gathered seismic data from offshore sedimentary basin settings as a testbed environment. Historical seismic survey datasets from Atlantic sedimentary basins are often supplemented with alternative geophysical surveys and site-specific geological constraints. Combining seismic borehole and stratigraphic logs with regional seismic datasets provides a link between field-gathered seismic waveforms, stratigraphy and depth-dependent, small-scale fluctuations in seismic velocity. This in turn enables the creation of synthetic velocity models and seismograms with field-derived properties, centring the collation of data for real-world machine learning applications in the numerical domain. We aim to bring insights gained from training FNOs on a better understood seismic environment to volcanic contexts in future work.