The visual interpretation of geological thin section is a meticulous endeavor carried out by geoscientific specialists in order to ground truth log interpretation as well as guide the spatial distribution of properties required by reservoir simulation models. At the same time, the shortage of qualified personnel, the abundance of dormant core data and the requirements for increased reservoir model accuracy have created operational needs that human interpreters alone can hardly fulfill.
In this context, a method for AI-assisted thin section interpretation was developed, leveraging the latest advances in the field of deep learning to provide geologists with a comprehensive set of reservoir properties derived from rock images. While a significant part of the solution relies on the training of supervised convolutional neural networks, establishing consistent labeling procedure, enforcing geological rules, removing input and output image artifacts and close communication with subject matter experts were equally critical ingredients to a geologically-realistic prediction as well as supplementing a scarce amount of input training data.
The main outcome of this multi-step domain-knowledge and data science work not only led to an increase in the mean intersection-of-union metric but also to the assurance that fundamental geological principles were honored. In practice, the algorithm ensured that petrographic object detection was constrained by biostatistical population criteria as well as prohibit the occurrence of non-natural combination of nested framework grain.
The aforementioned enhancements were subsequentially implemented and deployed at company scale for ADNOC's specialists to carry out their geological interpretation through conventional web-browser applications.
