Predicting deep-water depositional-element types on sedimentary logs using machine-learning algorithms trained on geological analogs

The contrasting architecture and heterogeneity of depositional sub-environments of deep-water sandstones has a large influence on reservoir geometry and properties. Thus, identification of the depositional setting of sandbodies is an important task in the interpretation of sedimentary successions and is routinely undertaken via the utilization of geological analogs or facies models. A workflow is proposed to automate this task by using machine learning models leveraging large quantitative analog datasets.

In this study quantitative facies and bedding data from 205 sedimentary logs of deep-water channel elements (104 instances) and terminal deposits (101) derived from 25 deep-water systems, stored in a relational database, were utilized to develop machine learning models trained to classify a facies sequence from a sedimentary log into one of these two element types. The classification is based on 18 quantitative facies and bedding attributes (e.g. thickness and lithology proportions). Given that the number of attributes interpretable from common wireline-log suites is limited, a ‘borehole’ attribute set was also used in model training, which only includes 5 attributes (sandstone and mudstone thickness and proportion).

Cross-validation results indicate that both sets of models identify element types correctly in more than 75% of the cases. Model performance in the application to an unseen testing dataset resulted in correct identification of 88% of the terminal deposits and 58% of the channel elements, with 69% mean accuracy overall. By contrast, models trained on the ‘borehole’ set of attributes only yielded a 64% mean accuracy. Model performance was compared to interpretations made by 14 sedimentologists on 12 graphical sedimentary logs selected from the testing dataset. Classifications made by the models (trained on the ‘outcrop’ attribute set), outperformed the geologists by 5% in accuracy. These results indicate that, to some extent, machine learning models are capable of assisting sedimentological interpretations of deep-water depositional environments.