Detection of framboidal pyrite size distributions using convolutional neural networks

Abstract

Pyrite (FeS2) framboids, spheroidal groups of discrete equant pyrite microcrysts, are found in sediments of all geological ages. The size of a pyrite framboid is established during early diagenesis and preserved through time. Framboid size distributions are hence useful for the evaluation of depositional conditions. In this work, we present machine learning approaches to characterize the size distributions of pyrite framboids to understand the intensity and duration of anoxia and euxinia during the Middle Devonian of the Appalachian foreland basin by analyzing framboid size distributions of the Marcellus Shale from Lycoming County, Pennsylvania. Importantly, we overcome the time-consuming and laborious nature of current manual tracing methods to enable the processing of high volumes of micrograph data. Specifically, we implement convolutional neural networks (CNNs) to characterize framboids from 14 samples across depths in the Marcellus Shale. We show that CNNs enable the precise and fast measurement of framboid size distributions from scanning electron micrographs. CNN architectures including Inception, ResNet, Inception-Resnet, and Mask R-CNN were trained and tested on a total of ~6,800 framboids from 128 grayscale and 32 colored scanning electron micrographs. Kolmogorov-Smirnov tests on the framboidal equivalent diameter distributions measured from CNNs and manual tracing show that the CNN algorithms detected framboids with up to 99% precision. Importantly, once trained, the CNNs were ~100 times faster than current manual tracing. A straightforward extension of this work includes the application of CNNs to characterize pores, fractures, organic matter, and/or mineral grains in geological materials.