GMIG has established itself at the forefront of research into deep learning and wave-based imaging, creating opportunities for innovative partnerships with industry collaborators. The group has developed a physics-based deep learning approach to wave-based imaging problems, introducing a new architecture called FIONet. Traditionally, through scales, imaging reflectors and smooth wave speeds have been treated separately, but in deep learning, this is accomplished jointly.
Working with Fourier integral operators (FIOs), GMIG has modeled and analyzed a wide range of wave-based imaging modalities in the past. FIOs now have a counterpart in deep learning, enabling data-driven, learned imaging. GMIG’s architecture exploits the geometry of wave propagation, turning the dyadic parabolic decomposition of phase space into convolutional layers and letting data determine the geometry via optimal transport and a critic network.
Deep learning and wave-based imaging have been integral components of the GMIG research program. The group pays particular attention to inductive bias, as a trained FIONet needs to perform in geological environments it has not seen before. Successes in wave-based imaging are particularly relevant to the group’s seismology portfolio, feeding into GMIG’s research on carbon sequestration monitoring alongside industry partners and peer institutions.
Further reading: Geometry of Wave-Based Imaging