Quantifying Uncertainty in Machine Learning Predictions for Seismic Reservoir Characterization

This paper explores techniques for quantifying uncertainty in machine learning predictions for seismic reservoir characterization, with a specific application to the Volve dataset, a deepwater oil and gas field in the North Sea. It distinguishes between two types of uncertainty: aleatoric, which comes from inherent data variability, and epistemic, which arises from incomplete knowledge or model limitations. To address epistemic uncertainty, the study applies ensemble-based strategies, specifically jackknife validation, to assess prediction variability. By incorporating synthetic data and transfer learning with real seismic data, the model’s robustness and ability to generalize are enhanced. The methodology provides valuable insights into areas of high uncertainty, improving the reliability of machine learning models for predicting reservoir properties in complex geological settings like Volve.