Advancing Reservoir Characterization with Machine Learning: A Multi-Well Predictive Analysis

Abstract Reservoir characterization remains one of the most significant requirements in optimizing hydrocarbon recovery, yet traditional methods still struggle with data variability, complexity, and cross-well generalizability. This study presents a supervised machine learning (ML) approach to predict production profiles across three wells, focusing on dataset structure variabilities, specifically the use of the whole dataset compared to perforated-zone subsets, impact model performance, and predictive accuracy. In this study, 926 data points from two training wells (A and B) was used for training and validations, while 917 points from an additional well C was used for out-of-distribution testing. Advanced data processing techniques, including variance inflation factor (VIF) for mitigating multicollinearity and singular value decomposition (SVD) for identifying hidden correlations across the three wells, were used. Ten different ML models were trained via randomized search optimization, with the Light Gradient Boosting (LightGBM) model achieving the highest accuracy (MAE: 0.0679, R2: 0.88). Testing on Well C revealed a deteriorated performance (MAE: 1.1907, R2: 0.66) and poor generalizability, especially for variables out of range compared to the training set. This can be attributed to the inherent geological variability and differences. The predictions deteriorate even further when focusing on perforated zones from the three wells (MAE: 0.866, R2: 0.459), indicating that additional factors may require further investigation to enhance prediction accuracy in these specific reservoir intervals.