Physics informed neural networks for velocity inversion

Abstract

In this abstract, a new neural network, physics-informed neural networks (PINNs) (M. Raissi, 2019) are introduced and implemented to solve the inversion problems of wave equations. PINNs employ standard feedforward neural networks (NNs) with the partial differential equations (PDEs) explicitly encoded into the NN using automatic differentiation, while the sum of the mean-squared error in initial/boundary conditions is minimized with respect to the NN parameters. Specifically, here we use this network structure to produce an accurate velocity model from seismic data. Our approach relies on training deep neural networks that are extended to encode the acoustic wave equation. In the first case, given analytical solution to an initial boundary value problem (IBVP) to infer very accurately the velocity parameter. In the second case, given seismic wavefield data in space-time, we use several coupled deep neural networks to infer vary accuracy the velocity field. After compared the results with full waveform inversion (FWI), the promising results for synthetic 2D data demonstrate a new way of using seismic data to identify key structures in the subsurface from machine learning approaches. Presentation Date: Wednesday, September 18, 2019 Session Start Time: 9:20 AM Presentation Time: 9:45 AM Location: Poster Station 2 Presentation Type: Poster