Researchers from Texas A&M University have developed neural network models that can predict the behavior of carbon dioxide in fractions of a second after its injection into deep saline aquifers. This is especially important for projects involving long-term CO₂ storage. Underground, the gas does not remain in place, but moves slowly, changing its phase state and partially dissolving in formation water, as well as getting fixed in the pores of the rock. In order for these storages to be safe and effective, one needs to understand in advance how far the CO₂ will spread and how much of it will be reliably fixed in the formation.
The biggest challenge is that CO₂ migration develops over decades and depends on many geological and operational factors, such as rock permeability and porosity, formation pressure and temperature, stratification, water salinity, injection speed and mode, etc. While conventional computer models give accurate results, they tend to work very slowly: one scenario can take up to an hour of calculation time. This makes it difficult to analyze a large number of variables, which is required in real projects with constantly changing operating conditions.
To speed up these efforts, the researchers created two different neural network architectures. The first architecture, a hybrid model combining long short-term memory (LSTM) and multilayer perceptron (MLP), predicts the geometry of the underground CO₂ plume, i.e., its expansion and movement in space over time. It generates a full forecast for 80 years ahead. The second model, built on the principle of sequence-to-sequence learning (Seq2Seq), solves another problem, i.e., it predicts the phase state of CO₂: what part of it will remain mobile, what will be permanently trapped in the pores and what will dissolve in saline water. Different tasks require different approaches and give different units of measurement accuracy: the first model works with distances in meters, while the second one deals with fractions of substances.
In order to train both models, the researchers conducted 773 detailed simulations, varying nine geological parameters and using 400 different CO₂ injection schedules realistic to the maximum extent, including pauses, flow fluctuations and other operational features. This database took about a month of calculations to create, but it allowed the neural networks to reliably reproduce the key features of gas behavior.
The tests showed high accuracy. The model predicting the development of the underground CO₂ plume gave an average error of about 42 meters. For objects where the diameter of the plume averages hundreds of meters and can reach one kilometer, this is a very small spread: even physical simulations often give errors on a comparable scale. The second model, which estimates the distribution of CO₂ phases, showed an error of less than 2% of the gas share, i.e., its predictions fully align with the standard curves, and the deviations are included in the calculated percentages of the full range of possible values.
Speed proved to be the main advantage. Where the conventional simulation required up to 60 minutes of calculations, the newly-trained models gave results in a fraction of a second. It is precisely this speed that allows the engineers to analyze dozens and hundreds of variations essentially in real time: changing injection schedules, taking into account possible well stoppages and checking the behavior of gas with different properties of the formation. This, in turn, makes the operation of underground CO₂ storages more predictable, manageable and reliable with long-term planning.
