Mineralogical Estimation of Organic Rich Mudrocks From Well Logs Using Neural Networks: Overcoming Training Dataset Size Limitation by Integrating X-ray Fluorescence Elemental Data
Mineralogical composition of rocks is one of the fundamental information that is useful in different disciplines in the oil and gas industry. For example, geologists use mineralogical composition in provenance analysis, geophysicists build rock physics template with specific rock composition range, and engineers use clay proportion to determine the optimal drilling and completion parameters. Traditionally, mineralogical composition is estimated by 1) petrographic analysis such as point counting or infra-red spectroscopy, 2) core examination, and/or 3) well-log analysis such as multi-min models. The success of these methods is variable and is highly dependent on the rocks examined. Organic-rich mudrocks mineralogical composition is harder to identify using these traditional methods because of their inherent 1) small grain size, and 2) highly variable nature at different scales. X-ray diffraction can be used but it is relatively slow and expensive. Neural networks can be used but they require a relatively large training dataset. In this work, we present a workflow to obtain an accurate mineralogical estimation by integrating relatively cheap and fast to obtain x-ray fluorescence elemental data (XRF) and traditional well logs. XRF data is inverted to mineral proportions using constrained optimization based on the stoichiometry of the expected minerals. The relatively large dataset obtained from the analysis can then used as training set to construct a neural network model with well logs as input and mineralogical proportions as output. Finally, mineralogical proportion is predicted with the neural network using well logs in intervals where XRF is not available. The workflow is validated using x-ray diffraction mineralogical data and illustrated using a real-world case study. The studied formation is the Shublik Formation, North Slope Alaska, where the rocks have highly variable proportions of calcite, quarts, illite, apatite and pyrite. Inverted mineralogy shows good correlation with independently the measured mineralogy from x-ray diffraction. Source code is provided for reuse. Generally, the integration of traditional analysis methods is essential to overcoming the limitations of machine learning methods in geoscience.
AAPG Datapages/Search and Discovery Article #90350 © 2019 AAPG Annual Convention and Exhibition, San Antonio, Texas, May 19-22, 2019