Geochemistry

Random forest

Geotechnical engineering

Mineralogy

Operating system

Well logging

Computer science

Thin section

Geology

Porosity

Section (typography)

Petrography

Artificial intelligence

Geophysics

Artificial neural network

Convolutional neural network

Diagenesis

Journal of Petroleum Science and Engineering

Images represent a large and efficient source of geological information from oil exploration. To better analyze them, well-known machine learning algorithms are used to extract mineralogy and porosity data from petrographic thin section images. Microscopic petrographic analysis allows obtaining images from thin sections in the visible spectrum. They are used to evaluate depositional environments and diagenetic processes during the formation of sedimentary basins. However, that is an activity subjected to the petrographer's experience. Data from other sources, such as chemical microanalysis, provide quantitative information that might assist in petrographic evaluation, but they are expensive and time-consuming. The main objective is to create models that systematically interprets mineralogy and porosity from images acquired of optical microscopic analysis using machine learning algorithms, standardizing descriptions and reducing subjectivity and human errors during thin sections analysis. Image segmentation models are created with representative classes of the rocks' mineralogy and porosity. Datasets were selected from images originated from thin sections of carbonate rocks, which are prepared from sidewall core samples of oil wells, specifically from the pre-salt reservoirs of Santos Basin, on the southeast coast of Brazil. These models use discrete convolutional filters followed by artificial neural networks and random forest classifiers. A number of configurations were tested, using different convolutional filters and classifier's parameters. Five models were created: 1. mineralogical model using artificial neural network; 2. mineralogical model using random forest; 3. mineralogical model using random forest validated by chemical measurements; 4. porosity model using artificial neural networks; and 5. porosity model using random forest. They were evaluated through the use of 10-fold cross-validation tests and by correlation with chemical microanalysis. Correlation between the two technique's relative occurrences for each mineral phase shows a root mean square error of 8.99% and a coefficient of determination of 0.82. That demonstrates how well models can generalize.

Digital petrography: Mineralogy and porosity identification using machine learning algorithms in petrographic thin section images