Stepped machine learning for the development of mineral models: Concepts and applications in the pre-salt reservoir carbonate rocks

Hybrid mineral model integrating probabilistic and machine learning approaches for the Brazilian pre‐salt carbonate reservoirs

“…Advances in the geochemical tool (Radtke et al., 2012) and an increase in the number of chemical and mineral composition analyses allowed the creation of less empirical direct models due to more complex mathematical functions capable of generating models from databases with little or no human interference (Freedman et al., 2015). Recently, a mineral model was proposed for the Brazilian pre‐salt rocks using a database with more than a thousand rock samples through the sequential training of machine learning algorithms (de Oliveira et al., 2021). This model estimates the fractions of calcite, dolomite, quartz, detrital clay minerals, K‐feldspar, plagioclase and pyroxene.…”

“…This bias is exemplified in Table 2, which presents the number of samples with compositional and mineralogical analyses collected in each pre‐salt formation in the Santos Basin, used in de Oliveira et al. (2021). More than 97% of the samples were collected in the Barra Velha and Itapema Formations, mainly composed of carbonates.…”

“…Another widely used approach, here called the direct model, consists of calculating the mineral fractions exclusively from the concentrations of chemical elements in the rock matrix acquired by the geochemical tool (Anderson et al., 1988; Flaum & Pirie, 1981; Herron et al., 2014; Zhao et al., 2017). A third class, called the machine learning model, was recently introduced where mineral models are created through trained artificial intelligence algorithms using experimental chemical and mineral composition analysis in rock samples and applied to geochemical logs (de Oliveira et al., 2021).…”

“…These models are the same ones presented in de Oliveira et al. (2021), trained using chemical and mineral composition analysis in rock samples. Next, these mineral reference curves are used as input into a probabilistic model along with density, photoelectric factor, geochemical and nuclear magnetic resonance logs for being able to provide information related to minerals and rock fluids.…”

Depth-variant pore type modeling in a pre-salt carbonate field offshore Brazil

“…To better understand and characterize these reservoirs, several authors have published studies with different focuses, such as unsupervised seismic facies classification (Ferreira et al, 2019), supervised artificial neural networks to predict rock property parameters such as porosity and acoustic impedance (Clarke et al, 2021), stepped machine learning algorithm to create mineralogical models from geochemical and mineralogical data (de Oliveira et al, 2021), thin sections analysis in the nonreservoir section focusing the identification of distinct magnesian clays and the processes of preservation and transformation of these minerals (da Silva et al, 2021), and clay and water saturation volumes through a hybrid method which combines Nuclear Magnetic Resonance (NMR) and conventional logs (Castro and Lupinacci, 2022).…”