In recent years, machine learning has become an increasingly important tool for automatic seismic interpretation. Only a very small percent of the available seismic data is used, since the interpretation of this data is highly time-consuming. The aim of this research is to develop machine learning based workflows which can aid geologists with interpreting the subsurface and automatically extracting more information from seismic images.
In geophysics, the most used technique for portraying the subsurface of the earth is reflection seismology. In reflection seismology, sound waves are emitted from a source and travel through the earth. As sound waves are travelling through the earth, they will pass through several layers of rocks with different properties, such as the rock density or the specific rock’s ability to withstand compression. When a sound wave hits a boundary between rocks with different properties a fraction of the sound wave energy will be reflected towards the surface. These reflected waves are recorded on the surface and used to construct the seismic image. This technique is in many ways like how doctors obtain images of organs or bones in the human body with the use of ultrasound or how certain mammals such as whales and bats use echolocation to construct an “image” of their surroundings which they use to navigate and search for food. After the seismic images have been constructed a geologist is tasked with interpreting them to better understand the geology and the geological history of an area. This is an essential part of oil and gas exploration since this understanding can be used to identify possible oil and gas reservoirs.
The proposed doctoral work will focus on the application of machine learning techniques for automated seismic interpretation and modelling of complex geological structures and stratigraphies in seismic images. Reflection seismic is the most commonly used technique used to portray the earth’s subsurface, where seismic waves are emitted from a source and allowed to propagate through the earth. Parts of the emitted waves will be reflected, which can be measured and compiled into seismic images of the earth's interior. Geologists use these images to create maps of the subsurface to better understand the geology in an area.
Machine learning describes a set of algorithms that aims to improve its performance with respect to a task, without explicitly being programmed to do so, by creating a machine learning model and exposing it to data. The field of machine learning can be divided into two subfields, supervised and unsupervised learning. In supervised learning, the model learns the mapping between labelled input-output pairs of data, whereas in unsupervised learning the model finds structure in the data itself without the use of labels.
In recent years, machine learning methodologies have been successfully employed in seismic interpretation workflows. However, many of these methods are based on supervised machine learning methodologies, which requires the use of labels in form of manual interpretations of the subsurface. The disadvantage with manual interpretations is that they are both time-consuming and require extensive domain knowledge to make. Since this is an interpretation task, the labelled data will also be sensitive to bias. Thus, the main motivation and goal of this research are to develop automatic/semi-automatic interpretation tools, based on unsupervised/semi-supervised machine learning techniques, which do not require manually labelled data. This approach may mitigate some of the challenges present in the supervised approaches as outlined above.