Physics-guided unsupervised deep learning approach for the inversion of receiver functions

SUMMARY The converted wave technique, namely Receiver function (RF), has been routinely employed for estimating 1-D velocity models of the Earth's crust and mantle structures. Physics-driven methods such as inversions are employed to RF data to estimate velocity models. Despite their wide utility, the presence of dipping and anisotropic geological structures often complicates this process. To address these complexities, here we introduce a Physics-guided unsupervised deep learning approach for the inversion of RF data. An unsupervised deep learning approach together with implicit neural representations is developed here, for prediction of subsurface model parameters such as the layer thickness, S-wave velocity, anisotropy, trend, plunge, strike and dip without requiring any labelled data. In addition, we incorporate a dynamic layer setting criterion to automatically pick the optimal number of layers required to fit the data, which is found to be particularly effective for identifying pronounced discontinuities like the crust–mantle boundary. For determining the optimal model parameters, neural network output parameters (the model parameters) are used in forward modeling of the RFs. Inversion results from both synthetic and real field data from the Indian shield and Hi-CLIMB network suggest that the physics-guided unsupervised deep learning approach is effective in inversion tasks, particularly when dealing with dipping and anisotropic media.