Predictive analysis of total organic carbon (TOC) in shale targets: example from the Lower Cretaceous of the Austral Basin (Patagonia, Argentina) using machine learning on outcrop data

Abstract

The Río Mayer Formation (Lower Cretaceous) of the Austral Basin, Patagonia, is a key source rock for unconventional reservoirs. This study explores the potential of machine learning (ML) for predicting Total Organic Carbon (TOC) content using outcrop data, a novel approach compared to traditional subsurface data applications. Employing dimensional reduction techniques (PCA, T-SNE, UMAP), the analysis revealed clear clustering of high TOC values in feature space, supporting the feasibility of predictive modeling. Three ML models— Logistic Regression, Support Vector Classifier (SVC), and K-Nearest Neighbors (KNN)—were tested using a feature set derived from ANOVA F-Score rankings. Dimensionality reduction improved model performance, with SVC achieving the most robust results. Despite limited labeled samples, predictions across models were consistent, identifying a promising region for high TOC. The study highlights the importance of integrating geological variables and XRD data in TOC modeling and emphasizes the need for expanded datasets and additional sedimentary sections to enhance regional interpretations.