Machine Learning in Geophysical Processing and -Inversion

Marine seismic data contains signals that represent waveforms reflected from the subsurface. These waveforms contain information about the physical properties of the subsurface, such as the velocities of the elastic wave that propagates in the geological layers. By using the information from seismic velocities, the elastic waves can be migrated back to their original reflection point in the subsurface and thus get a structural image of the geology. Seismic data are therefore essential for understanding the geological history of the Norwegian continental shelf. The process of collecting and processing seismic data, before interpreting the geology, consists of many processing steps and is computationally demanding. Lundin is an active contributor to improving all stages of this process. In this project, the candidate has used breakthroughs in deep learning to improve some of the geophysical processing applications and addressed issues that arise when applying such data driven technology to seismic wavefield data. During the project period, the candidate has collaborated with other fellows, which resulted in co-authorship of two articles dealing with deep learning applied in seismic processing, more specifically for noise removal. In the main part of the research, where the candidate himself has had the main responsibility, the focus has been deep learning for seismic data reconstruction/interpolation where different techniques from digital super-resolution have been studied. The candidate has published two journal articles as the main author and one recently submitted as a second author. During the project period, methodological elements have been presented at two scientific conferences and internally at Lundin, as well as contributed to a conference article. The work with the method development in the recently submitted journal article was carried out in collaboration with another research fellow and commercial supplier, where elements and knowledge from the two previously published articles were used. The newly developed method is already in production at the supplier, and results from the project will be presented at "The Biennial Geophysical Seminar 2022". During the project period, the candidate completed a 2.5-month stay abroad at Heriot-Watt University in Edinburgh, where his stay was interrupted due to Covid-19. The research conducted at Heriot-Watt was aimed at problems related to the fact that many steps in seismic processing depend on the data being represented on a regular grid. Limitations in the physical collection process mean that the data do not have the required regularity. During her stay at Heriot-Watt, the candidate studied methodologies used for image filling, and expanded and formalized these techniques to fit multidimensional seismic data provided in a deep learning framework.

Prosjektet har bidratt til å identifisere områder innen seismiskprosessering hvor nevrale-konvolusjonsnettverk gir et forbedret resultat sammenlignet med eksiterende metodikk. Spesielt muligheter som er knyttet til utfordringer ved å gi økt seismisk oppløsning har vi identifisert tre ulike metoder, som hver har sitt anvendelses område. En utfordring ved bruk av maskin læring er å finne data til å trene opp modellene. I to av artiklene har vi utviklet metoder hvor det ikke er nødvendig med en ekstern kilde av treningsdata for å identifisere prediksjons modellen. I det siste arbeidet har det blitt illustrert hvordan de-migrasjon kan benyttes som for å generere trenings data. Metodikken har bidratt til økt oppløsning samt automatisering av deler av prosessen som tidligere hadde flere manuelle elementer.

In oil and gas exploration, artificially generated pressure waves are used to extract important information about the subsurface physical properties as well as the rocks geometrical appearance. In seismic inversion you formulate the forward problem, in order to make synthetic data, and try to fit this data with the observed data (field data). The closer synthetic data resembles field data, the closer we are to estimating the true model parameters. In this study we will investigate how we can use state of the art algorithms from both geoscience and datascience in order to predict a model which is closer to the true model. The proposed doctoral work will focus on combining methods used in geophysical processing and inversion with methods within the field of machine learning. The candidate will conduct research towards the use of numerical methods for optimizing and adapting methods and algorithms within these fields. The motivation and overall goal is to develop machine learning workflows in order to improve the efficiency of seismic processing and inversion processes. In addition, as high resolution seismic, such as Broadband seismic, has become more and more sought after in the industry, attention will be directed towards research within the use of non-conventional streamer configurations, such as variable depth streamers and special VSP setups, and how they relate to seismic resolution. Case studies on field data will be implemented in this thesis work.