Back in May, Matteo Ravasi gave an online talk at the Norwegian Visual Intelligence Center.

In case you missed it, here is the Youtube video

Title: Self-supervised learning in seismic data processing

Abstract

Deep learning has taken the field of geophysics by storm. However, after the initial excitement accompanied by a variety of successful applications in interpretative tasks, the subsequent wave of solutions for seismic processing and imaging has so far not delivered as intended. The main reasons behind this initial unsuccess are: i) the fact that processing and imaging are likely to be framed as regression (or domain translation) tasks, where signal preservation is a must; ii) the lack of trustworthy ‘noisy-clean’ training pairs, and; iii) the inability to explicitly take into account the underlying physical process associated to a given processing task. A paradigm shift is therefore required where reliance on training data is relaxed and the physical process is included as part of the learning algorithm. In this talk, I will discuss a number of applications that my group is currently developing combining so-called self-supervised learning with classical inverse problem theory to solve seismic processing tasks ranging from denoising to wavefield separation and interpolation to simultaneous source deblending.