4D CO₂ Plume Monitoring: Hyper-Specialization of Interactive Deep Learning Networks using Transfer Learning. AAPG Academy Webinar: Summary & Review

     This webinar was mainly about the applications of deep learning networks trained on seismic attribute data in order to model CO2 plumes in and beyond the reservoir formation. Monitoring data is used to train and update the model. In this case, the model was trained in the seismic response to gas, or the gas signature on seismic. A key goal of CO2 monitoring is the identification of a breached top seal, and the identification of gas signatures can be very helpful in this regard.

     I was hoping that a recording of the webinar would be available so I could add some slides to this post, but that does not seem to be the case. If it does become available, I will update this post in the future.

  

Equinor’s Sleipner Field is the Example for the Study 

     Equinor’s Sleipner Field in the North Sea was used for the study. It is a long-running CO2 sequestration project that injects CO2 into a deep Jurassic sandstone, above and laterally farther away from the zone where the natural gas and formation CO2 are produced. Every two years, a new seismic survey is shot with the same parameters as before. Top seal integrity is of utmost importance.

 

Data, Training, and Learning

     The data consists of three seismic attributes: amplitude, iso-frequency, and relative acoustic impedance. The amplitude data sets are from pre-injection to current. These attributes are used to train the model. There are three ways in which the model “learns.”

Supervised learning – Audio-Visual model using tuning labels

Reinforced learning – this happens while the network trains. It works on feedback and rewards. It allows the network to abandon false positives.

Transfer learning – this refers to reusing knowledge from a previous network in new data sets

     Networks are tuned when new data comes in. The process is very repeatable.

     First, an In-Situ Baseline Model was created from Pre-Injection data (1994). The network learns from the tuning labels that are placed on it. In this case, it is tuned to seismic response to gas, which refers both to injected CO2 and pre-existing natural gas in the rocks. The pre-injection model is then tuned with each arrival of new data, and the model is transferred to the new model. Tuning label placement is predicted by the model. It is important not to “overtrain” the model, which will introduce too much statistical bias. Thus, they utilized the three-epoch method to keep the statistical parameters unbiased. The goal is to get just the right amount of training and bias. The deep learning network can capture hidden structures and relationships. After training and tuning, a geometrical representation of the CO2 plume is generated. This is done every two years when a new seismic survey is conducted. The model can differentiate between the two dry gas types after it learns. In this case, the plume extends through time to the south and then to the north. In this case, some possible top seal breaching occurred, according to the model. The model was trained to distinguish CO2 from existing methane in the shallower reservoir through a technique called latent space drift. Some layering was observed in the plume, which is consistent with geology since sandstones are more porous to gas than the shaley layers in between.

Sequential domain adaptation (SDA) – the most effective training occurs when a network is exposed to information for the first time. Overfitting to a dataset introduces bias. SDA allows a network to continually learn, to evolve. Here, what the network is observing is physical, not statistical. Networks with high levels of user interaction and input can be used and transferred to new models with different data.

 

Q&A

Seismic data must be in a specific format. It was run on data sets up to 1.7 TB.

Importance of a pre-injection seismic study: A study should be started before injection to identify in-place reservoir fluids.

Statistical forcing or bias can affect modeling and should be addressed. He uses the three-epoch method to reduce bias. Bias can also be good in that it creates a geologic context = geologic bias according to geologists. Thus, you want some bias in certain directions, but not statistical bias.

Interactive deep learning – if it is there and can be labeled, it can be trained. Gas signatures are an example.

The plume signature is derived by letting the network show the plume.

Can one train all 1994-2008 to be used for 2010? Better to be trained one at a time. If trained together, the results would be messed up – conflicting and incomplete labels would confuse the network. One can do multiple data training selectively – takes out the tuning labels.

    

References:

 

4D CO₂ Plume Monitoring: Hyper-Specialization of Interactive Deep Learning Networks using Transfer Learning. AAPG Academy Webinar. November 19, 2025.