Cost-efficient proxy modeling with active learning for CO2/N2-enhanced gas recovery and CO2 storage

Abstract

This study introduces a combined active learning and deep learning approach to develop a proxy model for carbon dioxide/nitrogen-enhanced gas recovery efficiency. Constructing traditional proxy models requires extensive and expensive training data from numerous reservoir simulations. In contrast to the random selection-based sampling of the control group, active learning optimizes this by selectively acquiring training data from specific reservoir scenarios. Both approaches start with 150 training data points and expand by five after each training session until the model achieves a 0.75 coefficient of determination with the test dataset. Focusing on a carbon dioxide/nitrogen -enhanced gas recovery setup with one injection well and two production wells, the model uses three input parameters to predict 13 output parameters related to gas production/injection volumes and carbon dioxide storage. The results show that active learning significantly reduces the number of required simulations by 54, on average, compared to random selection, thus enhancing efficiency. With the same simulation costs, the active learning models outperform the random selection by achieving a 17.3% higher coefficient of determination and 31.8% lower mean squared error. This improvement allows active learning to more reliably optimize enhanced gas recovery scenarios, identifying the top 2.6% of the optimal solutions from a pool of 500 scenarios, demonstrating its effectiveness in boosting enhanced gas recovery efficiency.

The authors' abstract, as published at the source. Gas Science and Engineering, 2026 · DOI ↗