Machine learning of the chemical elements for enhanced interpretation of depositional environments: Upper Jurassic strata case study in central Saudi Arabia

Abstract

To explore how machine learning (ML) techniques can enhance the interpretation of depositional environments using chemical elemental data, this study analyzes data from 156 shallow marine strata samples. Several advanced machine learning (ML) techniques were integrated to predict lithofacies and stratigraphic units of shallow marine carbonate strata, focusing on the Arab-D reservoir equivalent outcrops in Central Saudi Arabia. Utilizing a detailed geochemical dataset comprising major, trace, and rare earth elements, we aimed to determine the feasibility of using chemical elemental data in sedimentary geology for predictive modeling. The studied strata are categorized into two lithological units based on depositional settings: one unit comprises rocks that dominantly deposited below the storm wave base with a single lithofacies association, whereas the other includes rocks dominantly deposited above the storm wave base, and consists of two lithofacies associations. The study demonstrates the efficacy of ML algorithms in distinguishing these two lithological units, with a minimum prediction accuracy of 76%. Four ML techniques, including K-nearest neighbors (KNN); random forest classifier (RFC); gradient boosting classifier (GBC); and multi-layer perceptron neural network (MLP-NN) were used, two of them (RFC and GBC) achieved 89% and 92% prediction accuracy. However, a slightly poorer performance was observed in predicting fine-scale lithofacies associations of these lithological units, especially for the lithofacies association representing a transitional interval between the two lithofacies units. Expanding the scope of the study, we tested the ML models against geochemically similar strata from the Oligocene shallow marine reef carbonates. These strata are globally comparable and provide an external validation for the predictive models developed from the Saudi Arabian dataset. GBC and RFC maintained a higher accuracy level than KNN and MLP, demonstrating their robustness and transferability across diverse geological settings. Based on these results, the study underscores the significance of specific oxides and elements as distinctive geochemical markers, elucidating the environmental conditions of the depositional environments and their deposits. This research advances the application of ML in interpreting the depositional environments of shallow marine strata using elemental data, as the majority of the existing researches focus on employing unsupervised machine learning for cluster analysis.