2019 AAPG Annual Convention and Exhibition:

Datapages, Inc.Print this page

Automated Interpretation of Depositional Environments Using Measured Stratigraphic Sections and Machine-Learning Models

Abstract

For sedimentary geologists, the most common method for characterizing stratigraphy is the measured stratigraphic section from a core or outcrop. Various quantifiable metrics (e.g., bed thickness, grain size) distinguish stratigraphic datasets from different environments of deposition (EOD), but the relationships between these parameters is oftentimes qualitative, leading to subjective, ambiguous interpretation of environment. With the emergence and approachability of machine-learning models, there is an opportunity to create a more objective workflow for EOD interpretation. We present a supervised machine-learning model that utilizes labeled stratigraphic section and core description data to predict EOD. This model can be used to classify stratigraphic sections from unknown EODs and assess uncertainty for a given interpretation.

To enable this automated classification, we compiled measured section data (N=302 sections, n=30,000 lithologic beds) from four submarine depositional environments (basin plain, fan/lobe, channel, and levee). We chose data from well-constrained, unambiguous examples where other contextual data is available to confirm the EOD interpretation. In addition to basic parameters (e.g., bed thickness, grain size, net-to-gross), we extracted additional features (e.g. inter-quartile range of bed thickness, standard deviation of grain size) from each section. Using this labeled dataset, we train a variety of supervised machine learning models to predict depositional environment. Our model predicts the correct EOD with 73% accuracy.

When looking at subsurface core data from an area with poor geological context (e.g., very low-resolution seismic data), it can be difficult to discern the difference between two environments (i.e. external levee vs. distal lobe, channel axis vs. lobe axis). In these cases, our current workflow predicts EOD for a given core, and allows the geologist to assess the likelihood of that interpretation utilizing the uncertainty given by the model.

As large, high-dimensional datasets become common, the need for automated data analysis and classification is increasingly important, and stratigraphic data is no exception. Machine learning provides a valuable tool to accomplish the sometimes-subjective task of EOD classification (particularly from subsurface core data). Our modeling approach enable data-driven interpretation of depositional environment, which informs reservoir-modeling and well-placement decisions.