Inversion with the Born approximation in a deep learning framework

Zhan Niu, Jian Sun and Daniel Trad

ABSTRACT

Least squares reverse time migration (LSRTM) is an important technique that is starting to be used in the industry. LSRTM is closely related to Full-waveform Inversion (FWI) but instead of seeking for an optimal velocity model, it searches for an optimal reflectivity. Machine learning, on the other hand, has gained attention in the geophysics community and has become one of the most booming subjects in computer science. Various tools and methodology have been developed in the last few years and geophysicists have been finding applications by using these tools to solve more efficiently or with better quality long standing processing, imaging and interpretation problems. In this report, we first introduce an implementation of the Born modelling using the recurrent neural network (RNN) and second, we perform an inversion of the model by training the RNN with generated data. The inversion process can be proven to be same as LSRTM. The performance of different optimizers is compared and discussed. We conclude that the ADAM optimizer is the most stable and time efficient for this method.

View full article as PDF (0.91 Mb)