2019 AAPG Annual Convention and Exhibition:

Datapages, Inc.Print this page

Automated Grain Tracing and Point Counting Using Machine Learning

Abstract

Thin sections provide geoscientists with a wealth of information about a rock’s makeup and diagenetic history. For example, the amount of clay minerals or percentage of porosity can play a large role in the quality of a reservoir. However, the analysis of thin sections often requires many hours of manual labor, limiting the amount of analysis a single person can accomplish in a reasonable time frame.

Supervised machine learning brings the promise of automating time-consuming tasks, such as point counting and segmentation (i.e., identifying each pixel in a micrograph), to thin section analysis. In supervised machine learning, labeled examples are provided for the machine to learn from. Previous attempts using machine learning required an expert to hand design “features” to serve as inputs into a machine learning algorithm. These features could be mathematical representations of a grain’s characteristics such as cleavage, color, or twinning. The design of such features can be quite arbitrary, and features that work in one thin section may not work for others. Here we apply a recent development in machine learning that only requires traced grains as the input. The traced grains form a six-dimensional image (RGB color channels for both plane- and cross-polarized light images), and the resulting output is a fully segmented image. Preliminary results are promising, and may be comparable to point counting. Although the initial training examples still require human input, the model learned from the algorithm can be applied to other images that the algorithm has not been trained with. A fully segmented thin section can then be used to describe the morphology of grains (e.g., angularity, ellipticity), or serve as the basis for digital rock physics. This methodology can reduce time spent on a labor-intensive task and allow a user to move in to the interpretation phase more quickly.