Using ANN Prediction in Carbonate Reservoir Properties: Implication for Large-Scale Reservoir Correlation

Elemental data are typically presented and analyzed as ratios, indices, and proxy values. Due to the vast number of combinations, about 30-40% of the data often remains unutilized in correlations, yet this information could be key to comprehending geological processes. Automation and artificial intelligence (AI) considered as promising approaches to address This issue. This research aims to utilize and evaluate the performance of several Artificial Neural Networks (ANNs) architectures to predict lithological units (members) of a Permian-Triassic carbonate succession using integration of downhole elemental composition (ECS), mineralogy and other well-log data including delta-time (DT), neutron porosity (NPHI), density and gamma ray (GR). Principal Component Analysis (PCA) describes how much weight each original variable has in the principal components, i.e., the contribution of each feature to a PC. To address the classification objective of this study, several ANN architectures were designed and evaluated. Arch-I is a straight feed-forward neural network consisting of an input layer that accepts 14 features including ECS, mineralogy, and well-log data. Two dense layers with 512 and 256 neurons were used, dropout regularization was applied after each dense layer to prevent overfitting. Arch-II included 1024, 512, 256, and 128 neurons in successive layers, and culminated in an output layer with 4 neurons. Arch-III integrates batch normalization and activation layers after the initial dense layer with 512 neurons.PCA shows that the input training data of major elements and well-logs (DT, NPHI, density, and GR) could be grouped into three components that capture more than 80% of the variance, leading to significant reduction of dimensionality with minimal information loss. Over 50% of the variance is explained by the first component (PC1). Among the input ECS, mineralogical, and well-log data, elemental and mineralogical data show high correlation coefficient. The precision, recall, and F1 score for each architecture reflected the obtained accuracy. Among the used architectures, the simplest model (Arch-I) steadily showed the best performance, balancing model complexity with predictive accuracy, which scored 0.85. Arch-II and III have also shown comparably less accuracy of 0.84 and 0.81, respectively. The study successfully demonstrates the potential of using ANN architectures in integrating well-log data for geological prediction and correlation purposes. This finding emphasizes the importance of model selection in geological applications, where simpler models can often yield effective results for reservoir prediction. The findings also underline the value of ANN as a cost-efficient tool for enhancing geological interpretations by potentially reducing the time and effort required for dense data acquisition.