Seismic Soundoff

Arnab Dhara discusses his paper, "Physics-guided deep autoencoder to overcome the need for a starting model in full-waveform inversion," in the June issue of The Leading Edge. In recent years, physics-driven machine learning applications have been proposed wherein physics is integrated into the data-driven model to improve the ability of the machine learning methods to generalize and potentially overcome gaps in the physical theories. Solving geophysical problems by using hybrid physics-based and data-driven solutions has the potential to address simplifications in the physical models as well as overcome shortcomings with training data sets. Ultimately, they may refine and improve our understanding of the physics underpinning data sets. In this conversation, Arnab proposes employing deep learning as a regularization in full-waveform inversion. He explains why physics-based solutions with machine learning are challenging to develop, how he made it possible to train the network without known answers, and why