-
oa A High-Resolution Deep Learning Model for CO2 Injection at the Otway Field Validation Site
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, World CCUS Conference 2025, Sep 2025, Volume 2025, p.1 - 5
Abstract
The GeoCquest Field Validation (GFV) experiment aims to test and refine the approaches developed for predicting plume migration and trapping using high-resolution simulations. However, the multi-physics problem, highly nonlinear governing equations, necessity for high spatial and fine temporal resolution, and inherent uncertainty in the subsurface leads to computationally expensive numerical simulations that can take up to a week to run using conventional methods. Therefore, a U-FNO machine learning model was developed to maintain the complexity and high resolution of the simulations while providing a faster alternative that was then used to estimate breakthrough times at the monitoring well and characterize the uncertainty associated with the experiment. This model can also be used to examine reservoir characteristics, history match to PNL data and optimize injection strategies for trapping. The model created results in over 80,000x speed up in simulation time. Our model achieved a plume error of 3.2% for the training dataset and 6% for the test set. We used the model to understand the probability of CO2 reaching the monitoring well. We found that by day 23 of injection, there was over an 80% probability that CO2 would reach the monitoring well.
