How deep learning is changing Geoscience microscopy

Geologists have always used a wide variety of microscopy and microanalysis tools for a broad range of rock and mineral characterisation. As datasets become larger and projects become grander in scale, we are increasingly seeing the use of machine learning techniques to streamline all kinds of data acquisition and processing. These techniques have often highlighted the inconsistent nature of data handling by the geoscience community. This has commonly resulted in having lots of data but not “big data”, precluding the use of modern data analysis.

In many ways this is not the fault of the geologist. The complex nature of rock samples, covering more orders of magnitude in scale and requiring a detailed understanding of texture and chemistry, goes far beyond that of other materials in the physical sciences. Combined with analytical systems that were often designed for other sciences, this often results in a personal approach in how to interrogate our samples.

Light microscopy in particular provides a fantastic example of the personal nature of geological sample interpretation. Petrography training, and the subsequent ability to identify and interpret the mineralogy of a thin section, is the epitome of a standard geoscience task that has proved exceedingly hard to automate. Much of this is due to the vast number of minerals with overlapping optical properties, of which we use a dynamic (rotating polarisation) understanding to interpret a thin section. This is a combination of a huge amount of information with which we train our human brains, and standard data processing, even standard machine learning techniques are simply not “smart enough” to replicate.

Right now we are seeing the rapid emergence of deep learning neural networks (DLNN) across a range of 2D and 3D applications. In light microscopy these DLNN models can segment petrography data in ways that have never been possible before, with clear separation of minerals with similar appearance, distinguishing grain boundaries from fractures/cleavage, and determining boundaries of low relief minerals. They also have advantages over traditional segmentation in terms of lithology classification, where different combinations/textures of even the same minerals can have geological meaning but were previously hard to computationally separate. By taking these tasks that are relatively simple for human brains, but have been hard to upscale to large datasets, we can start to make consistent approaches across the geoscience community.

In addition, we now have the capability to generate large amounts of petrography data with automated slide scanners, meaning both data acquisition and processing can happen at the speed and scale necessary to make automated “big data” petrography a real tool for the geologist. Both data storage and the training of DLNN models can now be moved online, which means not only greater access of these tools to our community, but the ability to contribute to, modify, and assess the quality of future workflows.