Machine learning applied to geochemical missing data of the Vaca Muerta unconventional reservoir: Dealing with small sample sizes and heterogeneity

Abstract

Geochemical data provide comprehensive insights into reservoir chemistry, commonly obtained from well logs or through the laboratory analysis of core or cuttings samples. Data may be missing from certain intervals due to mechanical failures or geological complexity. This study evaluates the efficacy of machine learning methods in predicting missing element values within the Vaca Muerta Formation (Tithonian) located in Argentina, utilising X-ray fluorescence (XRF) results from 14 wells. Two approaches were evaluated: 1) Imputation methods to complete missing elements across specific depth intervals and testing four algorithms: K-nearest neighbour (KNN), multiple imputations by chained equations (MICE), miceforest, and Multiple Imputation with Denoising Autoencoders (MIDAS); 2) Prediction methods for estimating trace elements based on major elements concentrations, using selected “training” wells and comparing four ensemble tree-based algorithms: random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and Histogram-based Gradient Boosting Regression Tree (HGBT). Model validation involved statistical metrics (NRMSE and R2) and qualitative assessments, including comparisons between observed and predicted elemental values and comparison of wireline gamma-ray logs versus synthetic gamma-ray (derived from K, U, and Th). Imputation metrics indicated that KNN and miceforest outperformed alternative algorithms (NRMSE <0.1), with miceforest performing best qualitatively. The prediction results indicated similar accuracy among all tested algorithms, with a mean R2 reaching 0.83 in one of the two tested wells. In contrast, greater variability was observed in the second well (RF: 0.20–0.97; XGBoost: 0.05–0.96; LightGBM: 0.2–0.92; HGBT: 0.09–0.92). RF and XGBoost provided superior qualitative results. Both imputation and prediction methods exhibited reduced accuracy for Fe, Mg, Mn, Na, and P, primarily due to variations in depositional environments and diagenetic processes among training wells. This study demonstrates that small datasets (132–526 samples) and geological heterogeneity significantly impact model accuracy, and highlights the importance of combining qualitative and statistical evaluations.