Enhancing geothermal petrography with convolutional neural networks

The qualitative assessment of a geothermal reservoir using petrography is often conducted during drilling to assess the permeability, porosity, and mineral geothermometry of the reservoir rocks especially when little is known about the subsurface during the exploration stage. The petrographic analysis includes visually estimating pore and vein fractions, identifying rock textures, describing the degree of alteration, recognizing the hydrothermal alteration minerals and indicated temperature, and noting the presence of shearing and various porosity types. This traditional method of visual estimation and assessment is prone to errors when averaging fields of view and can be labor-intensive, especially during time-sensitive drilling operations when a geologist must analyze hundreds of thin sections per well under a polarizing light microscope. In this study, indicated porosity levels were assigned to 103 geothermal core thin sections based on the grouping of the rock parameters as observed under a polarizing light microscope. To enhance the traditional visual assessment in petrography, this study trained and validated convolutional neural networks (CNNs) in the automatic rating of porosity based on these parameters and in the detection of epidote, a key production marker in high-temperature magmatic-intrusive geothermal systems. Photomicrographs of the geothermal well core thin sections were utilized as input data for training and validating the ResNet, AlexNet, and VGGNet architectures. The three CNN architectures achieved porosity classification precision ranging from 0.74 to 0.84, and epidote detection precision between 0.90 and 1.0 in plane-polarized light (PPL) photomicrographs. The results demonstrate that CNNs can significantly augment traditional petrography in evaluating geothermal well samples.