Using Machine Learning to Transform Segmentation in Raw Materials
Great technological progress has been made over the last 20 years in the development of pore-scale imaging and modelling to address challenges in the petroleum geosciences. One of the principal challenges has been that these techniques are challenging to scale and automate, usually because the continuous outputs of the imaging techniques in question have to be ultimately classified into discrete phases for subsequent analysis and interpretation. These image outputs carry a variety of artifacts and noise that cause traditional analytical techniques to fail as the images become more complex. During visually examination, the brain of a trained petrographer, petrophysicist or mineralogist acts to integrate the rich, potentially multimodal datasets to extract the desired information. Such an approach is challenging to capture and express in a computational form, making microscopy challenging and expensive to scale across the many 1,000s of feet of core required to effectively describe reservoir behavior. Machine learning techniques give us, for the first time, a powerful set of tools to capture the complex set of processes involved in analyzing the rich datasets available to microscopic imaging in a way computational scalable to a much large range of samples.
In this study we will show, with the use of quantitative performance metrics, how such machine learning techniques perform when compared to more traditional image processing, segmentation and analysis techniques for a suite of different images, including both X-ray microscopy and nano-scale FIB-SEM imaging across a wide range of different noise levels. We will even show how such machine learning can be used to discriminate features which have little or no difference in their greyscale values, but instead are discriminated by textural features alone. We will also review a range of different applications of machine learning technologies to geological microstructural examination. First, we will show how it can be used to classify micro-CT volumes into different lithological regions, which are then used as a macroscopic map of geological heterogeneity, making resulting petrophysical pore-scale simulations more predictive of core scale behavior. We will then show how we can use machine learning to reduce acquisition time (and so cost-per-sample) for such petrophysical analysis. Machine learning based classifiers are much more noise tolerant than their traditional counterparts, and single high fidelity datasets can be used to train classifiers operating across a wide range of core sample. Finally, we will show how, by integrating automated mineralogical analysis with optical petrography, we can automatically extract mineralogical information from traditional cross-polarized light microscopy techniques. This is particularly exciting, as optical petrographic techniques are cheap, provide rich data about a wide range of reservoir petrophysics and geology, and can easily be scaled across extended sections of core.