CorePy: Visualizing, integrating, and upscaling core-based geochemistry through neural network-derived chemofacies models
Toti E. Larson, Esben Pedersen, Priyanka Periwal, and J. Evan Sivil
Core-based characterizations and analytical measurements provide critical information that is used to inform subsurface geologic models, ground-truth well-based wireline measurements, and allows geologists to understand complexities and heterogeneities in the rock. In our work flow, visual core-based lithologic descriptions at 1’ resolution are augmented with higher resolution (2” spacing) X-ray fluorescence (XRF) chemofacies core scanning. Rock and fluid attributes (i.e., mineralogy, oil and water saturation, rock mechanical properties, porosity and permeability measurements, SEM-based pore network distributions and lithologies, and organic matter characterization) are measured and/or imaged at specific core locations defined by XRF chemofacies characterization. Integrating all of these spatially disparate measurements, observations, learnings, and characterizations into a unified geologic framework that can be applied to all cores across a formation of interest has always been a challenge. Calibrating wireline logs from core-based measurements has been a particular challenge due to the typically low spatial resolution of key rock and fluid attributes measurements relative to the high spatial facies variability in typical mudrock systems. All of these challenges prevent the geologist’s ability to fully apply the findings from one or multiple cores to a project goal.
CorePy is a series of Python scripts that integrates high resolution core-based X-ray fluorescence measurements into a formation-specific chemofacies model using supervised training datasets that are hosted by the Mudrock Systems Research Laboratory (MSRL) consortium at the University of Texas at Austin Bureau of Economic Geology. A deep neural network model is trained on this formation-specific training dataset to classify the statistical probability that XRF measurements belong to any one chemofacies across all cores collected within a specific formation. Measured rock and fluid attributes are integrated into the chemofacies model to statistically characterize key attributes of each chemofacies. The chemofacies stacking pattern within each core is then used to model to mineralogy, porosity, and water saturation curves that are tied to wireline logs for further wireline log supervised training.
An important part of the CorePy work flow is validation. All chemofacies results are output and visualized directly onto core box photographs which puts the machine learning results onto a visual format that geologists can use to make core-based decisions. With this, changes to the training dataset can be evaluated and propagated to all XRF measurements in the database for unified consistency without delay time of visiting core warehouses. A script for image segmentation is added that masks fluorescing layers from UV photographs directly onto visible light photographs so that thin-bed volcanic ash layers can be highlighted and compared to chemofacies stratigraphic distributions. Future work adding image segmentation that can quantify laminated bedding is being pursued.
CorePy has been tested across mudrock plays in the United States including the Eagle Ford Shale, Austin Chalk, and Wolfcamp and Bone Spring Formations in the Permian Basin, with new projects in the Barnett Shale, the Marcellus Formation, the Bakken, and the Woodford Shale. In Argentina, we have developed a core-based chemofacies trained dataset for the Vaca Muerta Formation.
