Salt Body Flooding Using Activation Functions From Machine Learning

Abstract

In salt affected regions, conventional full-waveform inversion (FWI) is doomed to fail if there is no prior information of the salt body. Recent studies suggested to regularize the inversion by implementing an automatic flooding using total variation (TV) and Hinge loss functions. We generlize this approach and introduce a family of functions known as "activation functions" in the machine learning discipline that can be used to implement automatic flooding in similar way. In particular, we investigate the automatic flooding using a sigmoid, tanh and exponential linear unit (Elu) functions and apply them for salt body reconstruction on the BP model and report their performance.

Deep learning unflooding for robust subsalt waveform inversion

“…Therefore, the traditional practice aims to include the salt boundaries (ToS and BoS) in the initial model by means of the top‐to‐bottom workflow, and then follow that with low frequencies and long‐offsets FWI inversion. A more advanced approach is based on the fact that FWI can capture the ToS, which can allow for automatic flooding (Esser et al ., 2016; Kalita et al ., 2019; Alali et al ., 2020). However, these approaches cannot detect the BoS if frequencies are not low enough and offsets are not long enough, and, thus, still require human intervention to control the amount of flooding.…”