Paper Summary

Machine learning, and the variations thereof, have been around for a considerable time, perhaps as far back as the early 19th century with Bayes work on probability theory. It is then a bit of a surprise that these techniques have not made more headway in seismic processing, a data rich industry that should naturally suit greater automation using machine learning algorithms. Mathematical functions used in seismic processing are highly evolved and designed to improve the data quality by removing noise, correctly positioning data, or enhancing the data quality. The decades long development of these applications means they are extremely good at solving the challenges each one individually addresses. So how can machine learning help when the industry already has highly evolved and effective tools? There are two possible categories for applications, one to supplement the existing functions in a framework that enables greater autonomy, and the second to replace tools that are not as effective as they could be, allowing a greater diversity of testing-free applications (generalization). In both cases, the goal is improved data quality, faster. We present examples that might benefit seismic processing, as well as comment on the challenges faced with these methods.