High Throughput Mineral Classification Using Machine Learning
Dr. Matthew Andrew
Geoscience & Oil and Gas Technologist
Automated Petrography - High Throughput Mineral Classification Using Machine Learning
One of the biggest challenges in the microstructural characterization of geological media is that of scale. Frequently the resolution to describe fundamental rock structures comes at the expense of a field of view representative of geological heterogeneity. Scanning electron microscopy (SEM)-based automated quantitative mineralogy (AQM) techniques have developed to be a primary petrological characterization techniques, and while they can reveal significant insights on a specific location or sample, they can frequently be challenging or expensive to scale over the dozens, hundreds, or thousands of samples required for basin, formation, or field-scale geological characterization campaigns. Optical petrography has remained the mainstay for such large scale analyses, requiring the manual measurement and quantification of mineral, crystal and grain structures indicative of geological processes. This has historically required the exacting, arduous work of trained petrographers, a process which has proved extremely difficult to scale using traditional computational techniques coupled with automated imaging.
The last 20 years have seen a transformation in a broad set of applied statistical techniques broadly classified under the umbrella of “machine learning.” While these technologies have transformed areas as diverse as stock market analysis and medical diagnosis, its application to microscopy in general and geological microanalysis in particular has been limited. In this webinar we will review recent developments in automated geological microanalysis, coupling automated multi-polarized slide handling and image acquisition with advanced image processing and machine learning based pixel classification. This allows for mineral classification to be performed directly from the digital light microscopy data, which can then be streamed automatically to powerful image analysis tools allowing for grain sizes and shape, mineral associations to be measured, as well as sample wide modal mineralogies. These machine learning models can be either trained manually or can be correlated with SEM based AQM to allow for automated mineral training. As analyses can be performed much more rapidly using optical petrography than with SEM based techniques, locally trained models can then be scaled across many samples allowing, for the first time, large scale campaign analyses to be automated.