Deep learning in time-lapse processing (Alali et al., 2020)

Time-lapse (TL) seismic, also known as 4D seismic technology, is an essential tool to enhance hydrocarbon recovery by monitoring fluid substitutions in a reservoir. It also plays a major role in monitoring CO2 injection in carbon sequestration projects. In TL, multiple seismic surveys are performed with the same acquisition parameters. The quality of the TL data depends on the repeatable signal (4D signal) relative to the non-repeatable noise (4D noise). With ideal repeatability, the TL data from different vintages should be identical except at the target area. Unfortunately, TL data always have non-repeatable signals due to many factors such as ambient noise, velocity variation in the near-surface, source-receiver positioning, and other survey factors. 

Traditionally, 4D processing is implemented by matching the monitor to the baseline data through cross-equalization techniques (XEQ) such as matched-filtering. XEQ is implemented by finding a convolutional operator matching the monitor to the baseline data.  To avoid over-matching that can zero-out the differences caused by the target signals, the operator is found in a shallow time window such that it does not contain any reservoir signal. This setup does not allow us to accommodate variations of the needed correction with time.

Here, we test two deep learning (DL) models to implement the matching between the base and the monitor. Similar to the conventional XEQ, we propose to train the DL models on short shallow windows that do not contain the target signal and then we infer to the deeper parts. This allows the network to learn the static variation needed to match the surveys while preserving the reservoir signal.

 


DL models

1- Convolutional Autoencoder (CAE) + Fully connected neural network (FCNN)

We train a convolutional autoencoder (CAE) and a fully connected neural network (FCNN), separately. The goal for the CAE is to learn a representation of the data in a reduced dimension since it is easier to match in a smaller space. The FCNN model is responsible to implement the matching between the monitor and the baseline.

 

 


2- Recurrent neural networks RNNs

RNNs are used widely for time series modeling and predictions due to their history (memory) feature. Seismic data are time series; therefore, we propose to use RNN to map traces from the monitor to the baseline

References

Alali, A., Kazei V., Altaf B., Zhang X., and Alkhalifah T., (2020), "Time-lapse cross-equalization by deep learning", 82nd EAGE Conference and Exhibition 2020, European Association of Geoscientists & Engineers.

Alali, A., Kazei V., Sun B., Smith R., Nivlet P., Bakulin A., and Alkhalifah T., (2020), "Cross-equalization of time-lapse seismic data using recurrent neural networks", SEG Technical Program Expanded Abstracts 2020, Society of Exploration Geophysicists.

Alali, A., Kazei V., Altaf B., Zhang X., and Alkhalifah T., (2020), "Time-lapse cross-equalization by deep learning", 82nd EAGE Conference and Exhibition 2020, European Association of Geoscientists & Engineers.

Alali, A., Kazei V., Sun B., Smith R., Nivlet P., Bakulin A., and Alkhalifah T., (2020), "Cross-equalization of time-lapse seismic data using recurrent neural networks", SEG Technical Program Expanded Abstracts 2020, Society of Exploration Geophysicists.