Learning to Invert Pseudo-Spectral Data for Seismic Waveform Inversion

Abstract: Full-waveform inversion (FWI) is a widely used technique in seismic processing to produce high resolution Earth models that fully explain the recorded seismic data. FWI is a local optimisation problem which aims to minimise in a least-squares sense the misfit between recorded and modelled data. The inversion process begins with a best-guess initial model which is iteratively improved using a sequence of linearised local inversions to solve a fully non-linear problem. Deep learning has gained widespread popularity in the new millennium. At the core of these tools are Neural Networks (NN), in particular Deep Neural Networks (DNN) are variants of these original NN algorithms with significantly more hidden layers, resulting in efficient learning of a non-linear functional between input and output pairs. The learning process within DNN involves iteratively updating network neuron weights to best approximate input-to-output mappings. There is clearly similarity between FWI and DNN. Both approaches attempt to solve for a non-linear mapping in an iterative sense, however they are fundamentally different in that the former is knowledge-driven whereas the latter is data-driven. This article proposes a novel approach which learns pseudo-spectral data-driven FWI. We test this methodology by training a DNN on 1D multi-layer, horizontally-isotropic data and then apply this to previously unseen data to infer the surface velocity. Results are compared against a synthetic model and successfulness and failures of this approach are hence identified.