• Search
  • LinkedIn
  • Instagram

Deep-learning possibilities

Words by Kyle Watts
1 September 2022

The subsurface sequestration of CO2 is recognised as a promising mechanism to reduce anthropogenic CO2 emissions in the atmosphere. The movement and fate of the injected plume of CO2 must be closely monitored to identify potential leakage zones and ensure the safe, long-term storage of CO2 in the subsurface. Harpreet Kaur, together with colleagues at the University of Texas, USA, and the China University of Geosciences, designed a deep-learning framework to aid effective subsurface monitoring using time-lapse seismic data.

Using a range of porosities and permeabilities for a selected reservoir, the team generated models of dynamic reservoir properties, such as saturation and velocity. They then trained the deep-learning model using several time-lapse seismic images and their corresponding CO2 saturation values at an injection site. During the training phase, the model successfully learnt to map changes in CO2 saturation from the time-lapse seismic response.  The trained model could then be applied to datasets consisting of different time-lapse seismic image slices (which correspond to different time intervals and that were generated using different porosity and permeability distributions) to estimate both the CO2 saturation values and the extent of the CO2 plume. The authors test different cases to verify the effectiveness of their method. 

The algorithm provides a deep-learning assisted framework for 4D seismic inversion that can directly estimate CO2 saturation and plume migration in heterogeneous formations. This method bypasses numerous intermediary steps that must otherwise be completed when using the conventional time-lapse inversion workflow. Additionally, this approach incorporates the geological uncertainty associated with a selected reservoir by accounting for the statistical distribution of reservoir properties, such as porosity and permeability, throughout the training phase – all without requiring any theoretical knowledge about the error distribution, which is typically difficult to establish. 

Kyle Watts

DETAILS: Interpretation 10(1) (2022); https://doi.org/10.1190/INT-2020-0205.1

Related articles