Scaling sedimentary geoscience interpretations using machine learning: examples from fluvial to deep marine strata

Abstract

Machine learning, cloud computing, and open-source programming languages have changed scientific research and discovery for many fields. This thesis explores applying some of these technologies and techniques to sedimentary geoscience problems, along with standard numerical geoscience techniques. We examine three different modeling scenarios: fluvial scaling relationships, well-log property prediction, and core imagery analysis and core classification. Each scenario requires different methodologies and showcases different needs and capabilities of numeric modeling and machine learning for sedimentary geoscience research problems. The first study explores outcrop and satellite scale (plan view) metrics to reconstruct paleoflow conditions of the Triassic Lithodendron Wash Bed in Petrified Forest National Park, Arizona. These two independent measurement methods have been developed though modern river observations. For a specific stretch of the Lithodendron Wash Bed, this study confirmed that the two independent measurements roughly are equivalent for predicting the overall scale of flow conditions of this specific fluvial system. The second scenario utilizes a subsurface well log and core dataset acquired for hydrocarbon exploration and development. This 12-well dataset provides a testbed to examining various machine-learning strategies to predict core-based reservoir properties. This scenario confirmed that gradient boosted decision trees are an effective supervised machine learning model for tabular geoscience data and hypothesized relative differences in basin history recorded in the geochemistry of sandstones and mudstones. The final study applies machine learning models to a curated core-tray dataset from the United Kingdom Continental Shelf. The work covers an end-to-end workflow for core-tray images to lithology and facies prediction of deep-water gravity flow deposits, with depth registered to the 0.5cm scale. The scientific findings, resulting models, and developed code will have future implications for machine learning and standard numerical modeling research for sedimentary systems. This thesis serves to be a guide – and hopefully a model – for future research in this area, with all data, code, and methods being fully released and open source. Lastly, the analysis also identifies drawbacks to using machine learning and emphasizes that not every geoscience project needs advanced machine learning methods when many times a more straightforward, traditional approach would suffice.