Inversão da forma de onda completa com abordagem multiescala usando redes neurais recorrentes

Abstract

Seen as a fundamental parameter for a reliable geological image of the subsurface, and consequent exploratory success, seismic velocity is one of the prerequisites in the seismic exploration chain. Such velocity relates directly to the construction quality of seismic images through robust algorithms such as reverse time migration (RTM) or other seismic imaging techniques. In this work, we use machine learning environments to obtain high resolution seismic velocity through the full waveform inversion (FWI) technique. In summary, the FWI technique aims to compare data from real observations with calculated data obtained through seismic modeling from the solution of a specific wave equation. The residual of this comparison is minimized and the gradient is used to update, with an iterative optimization algorithm, the velocity model that at the end of the process will be able to correspond to the real data. Here we will use a Recurrent Neural Network (RNN), based on the governing physics (acoustic wave equation), to derive the real data and the calculated data regarding the direct seismic modeling step, since we will deal with purely synthetic data. In addition, learning environments, such as Pytorch, provide us with tools for calculating the gradient (automatic differentiation) and the mini-batch strategy important in terms of reduction memory and higher processing velocity. As the FWI is based on the iterative minimization of a cost function between observed and calculated data, in order to avoid convergence to local minima, we use the multiscale approach of frequency. In addition, we tested the FWI response when subjected to noisy observation data and less accurate initial models, and compared it with the inversion response added to the multiscale approach of frequency technique, to demonstrate the performance of this approach when it comes to mitigating these limitations. The results obtained in three sets of data demonstrate the efficiency and applicability of the technique used in the attempt to obtain high resolution seismic velocity fields.