Uncertainty Estimation in Reservoir Property Prediction Using Machine Learning: A Gulf of Mexico Case Study

Abstract Accurate prediction of reservoir properties is essential for reducing exploration and development risks in subsurface resource evaluation. Machine learning (ML) techniques, particularly deep learning, have demonstrated strong capabilities in mapping reservoir properties from seismic data. However, the reliability of such predictions remains a critical concern, especially in data-scarce environments. This study addresses this challenge by focusing on uncertainty estimation in ML-based reservoir property prediction, using a case study from the Gulf of Mexico. A rock physics-guided convolutional neural network (CNN) approach was used, with the CNN trained on a synthetic well catalog. The catalog was generated by calibrating a rock physics model based on four real wells, with a fifth well reserved entirely for blind testing. Scenario modeling was then applied to capture a wide range of geological variations, including changes in reservoir thickness, porosity, and water saturation. For each pseudo-well, synthetic seismic gathers were simulated, resulting in a comprehensive training dataset for predicting P-impedance, porosity, and water saturation. To evaluate the model's robustness and quantify uncertainty, the jackknife validation technique was applied. In this approach, one real well was excluded from the training set in each iteration, enabling the creation of multiple model realizations. This allowed the computation of statistical measures, such as the mean and standard deviation of predictions, offering insights into prediction confidence. Following the initial training, transfer learning was used to fine-tune the CNN using real seismic data and four real wells. The CNN-predicted property volumes showed strong agreement with filtered log curves, including at a blind well, demonstrating the model's generalization capabilities. Uncertainty maps revealed that areas in the central and northwestern regions of the study zone exhibit high porosity and low water saturation with correspondingly low uncertainty, indicating high confidence in predictions. Conversely, areas with high uncertainty – particularly in the southeast – highlight possible influences of seismic noise or complex geology. This integrated approach, combining synthetic data generation, rock physics modeling, machine learning, transfer learning, and jackknife-based uncertainty quantification, offers a robust framework for reservoir characterization. It is especially valuable in regions with limited well control, providing not only accurate predictions but also actionable measures of confidence for decision-making.