An implementation of a recurrent neural network for 1D acoustic waveform inversion

Abstract

Recurrent neural network (RNN) is a class of artificial neural networks widely used to model a temporal dynamic system. Recently, recurrent neural networks have been developed for acoustic waveform modelling in 1D and 2D bounded domains. Since the trainable parameters of the networks are the acoustic wave velocity, the process of network training is equivalent to solving an inverse problem of acoustic waveform inversion. In this work, we extend the previously proposed RNNs for acoustic waveform modelling/inversion in 1D unbounded domains by incorporating perfectly matched layers (PML) into the RNN cell. The proposed RNN architecture was implemented using Tensor Flow and has been successfully tested on a 1D synthetic data set. The results show that we have successfully implemented PML to RNN base acoustic full waveform inversion.