Abstract
Understanding rock mineralogy is essential for formation evaluation, improving the calculation of porosity and hydrocarbon saturation. The primary method to obtain the mineralogy from a well is by applying a model to the geochemical tool's chemical elements. However, creating a mineralogical model presents challenges such as the minerals' chemical composition and the decision to include a mineral in the model. The traditional application of machine learning can make mineral models less realistic since conventional training is developed based on a set of minerals with different occurrences, lowering some minerals' representativeness. The present research proposes the stepped machine learning (SML), a stepped way to use machine learning to create a mineralogical model from chemical and mineralogical data. A database was assembled with the elemental concentration obtained with XRF analyses and the mineral concentrations obtained with XRD analyses. The chemical elements were Al, Ca, Fe, K, Mg, Mn, Na, Si, and Ti. The minerals were calcite, dolomite, quartz, clays, K-feldspar, plagioclase, and pyroxene. Four algorithms were tested: MLP, GAN, Random Forest, and XGBoost, with XGBoost showing the best results. SML was applied, where a mineral model results are used to train a subsequent model. SML allowed for a significant improvement in some models, notably to clays with an increase in R2 from 0.597 to 0.853, quartz an increase from 0.673 to 0.869, and calcite, from 0.758 to 0.862. A decrease in the mean squared error of these minerals' models was also observed. The model was applied to the geochemical logs from three wells drilled in the Brazilian pre-salt, and the results were compared with XRD analyzes. The SML model was able to honor the mineral concentrations for different rocks. It is demonstrated that the integration between machine learning tools and geological knowledge in SML was crucial for creating a representative mineralogical model.