Do you know what junior geologist Wouter Bell Gravendeel is doing in the image shown above? Together with his fellow colleagues Jostein Larsen Hesthammer and Jahn Brusdal Eriksen, he is drawing. Not an outcrop or a landscape, but outlining each grain in a cutting sample. Or, as Jahn describes it: “a million-piece puzzle”.
Why would they do that? The answer is the power of machine learning.
As part of a collaborative project between M Vest Energy and Glex, a team of people consisting of developers, physicists and geoscientists is working on making the enormous amount of data becoming available through the Released Wells Initiative (RWI) more accessible. And not only that: the idea is to use information on individual cutting pieces (grains) in combination with drilling parameters, well log and mineralogy data.
Again, why would you do that? “The main rationale is to enable geoscientists to decode the subsurface in the most efficient way,” says project leader Alexandre Verechtchaguine from M Vest Energy, who is responsible for developing the machine learning algorithms. “To do that, we aim to combine all available data, including images, but also well logs, geochemical data and drilling parameters.”
Attend DigEx 2022, 5-7 April in Stavanger, and hear all about this project from Alexandre Verechtchaguine – M Vest Energy.
To get the information about individual cutting pieces, one has to delineate each of them on a photo. This can be done by a process called “segmentation”, which is a hot topic in machine learning at the moment. However, before any algorithm can be let loose on the data, it must be trained. Therefore, it is very important to prepare a proper training dataset for the project.
Although it is time-consuming to manually draw each grain in a training image, they form the ultimate benchmark against which the results produced by the algorithm is being compared. And once trained, the algorithm should be able to segment (most of) the images in the RWI database.
“One of our goals is to use the segmentation process to predict mineralogy using machine learning. After the individual grains have been identified, they will be labelled with the help of the available mineralogical and chemical composition data, and the machine learning algorithms will be trained to predict the minerals for all cuttings samples also outside the training dataset,” further explains Alexandre.
However, already at the current stage the segmentation has high value: it provides detailed information about the shape of individual grains. “We should remember that in many reservoirs, the sizes, shapes and sorting of the grains are the main controlling factor for porosity and permeability,” adds Alexandre, who also has a background as a petrophysicist. “In addition, it is also good to be aware that the information about grain colours and UV-light images are also available to us already,” concludes Alexandre.
The shape of cuttings can be divided into two to three main categories: angularity (angular to well rounded), cleavability (blocky – splintery – fissile) and a degree of dissolution/structure. “At the same time, it is important to be aware that the size and shape of cuttings is not only determined by lithology of the formation drilled, but is also to an extent dependent on the drilling assembly and drilling practices,” says Brit Thyberg from Glex AS.
“Combined with other data categories, the shape of cuttings can tell us something about the geological and sedimentological history, optimal/suboptimal well design, choice of drilling assembly, drilling practice and choice of mud type,” says Anita Hansen, who is a geoscientist at Chrysaor. “All these aspects can then be used in future interdisciplinary projects within subsurface, drilling and production,” she adds, “so the Machine learning approach taken by M Vest Energy is very interesting”.
So, when one wants to use cutting material from several wells to make inferences on lithological trends, one ideally needs to normalise and integrate cutting grain shape data against the parameters described above. “In Glex Energy, we can host all these data types and perform these calculations,” Jørgen Engen Napstad says.
The team at Glex and M Vest Energy is still working on the algorithms, but a lot of progress has been made recently. And once in place, the benefits will be significant. Not only will machine learning be performing the task much quicker, but it will also be able to combine different datasets in a way that has never been done before.
HENK KOMBRINK
