FWI with Optimal Transport: a 3D Implementation and an Application on a Field Dataset

We present the application to a 3D real dataset of full waveform inversion (FWI) with optimal transport (OT) using the Kantorovich-Rubinstein (KR) distance as proposed by Métivier et al. (2016). This approach involves an efficient numerical implementation for OT in time and space directions, allowing taking into account lateral coherency of the traces; this has an important impact on the quality of the results. The approach also exhibits a slightly reduced sensitivity to local minima compared to least square (LSQ) misfit. Moreover the iterative method used for the computation of the KR distance allows producing a set of intermediary solutions that span progressively from LSQ to OT. We recall the main components of the approach and present its numerical implementation in 3D. We show the improvement of the results compared to conventional FWI on 2D synthetic and 3D real datasets for the same number of velocity update iterations.