Unsupervised and Supervised Machine Learning to Classify Oils from the Norwegian Continental Shelf North of 62°N

Summary Unsupervised and supervised machine learning algorithms were applied on two large geochemical data sets, one from the Norwegian Sea and an other from the Barents Sea. Data bases were thoughtfully examined with regard to sample type, contamination, and data integrity, respectively. The purpose of this work is to compare both traditional and machine learning approaches for oil-oil and oil-source rock correlation on two large datasets on the Norwegian Continental Shelf. Unsupervised clustering of oil samples produces classes that are, overall, consistent with geological understanding of the origin of the oils. We established a workflow using supervised learning that enables an expert to infer oil types and source rock correlation in an immature (low confidence) area given available samples in a mature (high confidence) area. We emphasize through examples the importance of data quality and data consistency when using machine learning for automatic oil typing.