Seismic inversion by multiscale Newtonian machine learning inversion

Abstract

We present a multiscale wave-equation inversion method that inverts low-dimensional representations of seismic data to reconstruct the subsurface velocity model. The low-dimensional representations of seismic traces consist of low-rank latent space (LS) features predicted by well-trained autoencoder neural networks. The LS features mainly contain the kinematic information of seismic data, such as traveltime, when the LS feature dimension is small. However, large-dimensional LS features can preserve more complex characteristics of the seismic data, such as waveform variations. Therefore different scales of the subsurface model can be recovered by inverting the LS features with different dimensions. We denote this approach as multiscale Newtonian machine learning (NML) because it inverts for the model parameters in a multiscale way by combining the forward and backward modeling of Newtonian wave propagation with the dimensional reduction capability of machine learning. Numerical results demonstrate the success of multiscale NML inversion in recovering both the low and high-wavenumber velocity information. The advantages of this method are: (1) The LS features are automatically generated by autoencoders where no traveltime picking is required, (2) it is less prone to getting stuck to local minima than full-waveform inversion (FWI), (3) limited storage space is required because only the low-dimensional representations of the seismic data (> 100 times smaller) are needed to be stored.