Proxy modeling for CO2 injection in tight gas condensate reservoirs using active learning-based artificial intelligence

This study proposes active learning-based artificial intelligence application to efficiently build a proxy model for carbon dioxide (CO2) injection scenarios in tight gas condensate reservoirs. In gas condensate reservoirs, as production progresses and reservoir pressure decreases, condensate accumulation in the reservoir pores leads to a relative reduction in gas permeability, thereby lowering gas productivity. Injecting CO2 into the target gas condensate reservoirs to maintain pressure can mitigate condensate banking while simultaneously enabling CO2 geological storage. However, multiple variables influence the performance of such CO2 injection strategies. In this research, proxy modeling for a tight gas condensate reservoir mimicking the Montney region in Canada was performed using active learning, which optimizes the acquirement of additional training data. The proxy model was constructed with a random forest algorithm trained on reservoir simulation results generated using Petrel, Eclipse, and MEPO software from SLB. Initially, simulations were conducted for a limited number of scenarios, and additional data were iteratively acquired by identifying input scenarios with high uncertainty in predictions from the previous proxy model. This active learning process improves the efficiency when adding extra training dataset, enhancing the model's performance while reducing the need for exhaustive simulations. The input parameters for CO2 injection included the timing of switching a production well to an injection well, the bottomhole pressure of an injection well, and the maximum production rate. Output parameters included CO2 molar injection and production rates, field gas and oil production totals, field oil saturation averages, field gas injection cumulative total, CO2 storage total, and field average pressure. Experiments analyzed the minimum additional data required to achieve an R2 score of 0.95, with initial datasets of 30, 40, 50, and 60 simulations. For these initial dataset sizes, the active learning method saved an average of 4, 6, 3, and 1 reservoir simulations, respectively. Considering that each reservoir simulation requires an average of 45 minutes, the computational cost savings are significant. This efficiency is expected to be even greater for more complex reservoir simulations.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.