Let’s start this story with a philosophical question: “Isn’t there a contradiction in using human interpretation to ground-truth a model if the purpose of the exercise is to identify leads and prospects that would have been missed by a human expert?”
“It is one of the fundamental questions one needs to be aware of when embarking on a project like this,” says Peder Aursand from Aker BP.
Peder is at the forefront of developing machine learning algorithms to ultimately tackle the holy grail of digitalization in exploration: being able to directly identify or high-grade leads and prospects using an artificial intelligence approach.
“What is needed is an accurate and balanced source of ground truth, positives and negatives, as well as access to large amounts of relevant and consistent input data such as seismic, well logs, and more,” Peder adds.
However, subsurface data collection is not done in a systematic way. This leads to a fundamental bias in areas where high-quality data is collected, something that will be directly reflected in any machine training data built on it.
Added to that is that the black box nature of machine learning means that once the issues of data and ground truth has been solved and a model has been trained, one will face challenges in how to integrate it into exploration work.
“There are both social and technical aspects to this, with topics like model trust, uncertainty quantification and transparency being key,” as Peder further outlines.
“However, the ability of machines to process and combine data sets does offer huge potential, so addressing the above-mentioned hurdles in a systematic way and being aware of the limitations justifies the effort,” Peder concludes. “I’m looking forward to share our latest insights with the audience at DIGEX next week.
HENK KOMBRINK
