Machine Learning for Reservoir Characterization and Time-Lapse Seismic Analyses



Journal Title

Journal ISSN

Volume Title



Successful reservoir management requires a clear understanding of reservoir distribution, quality, heterogeneity, and baffles. During field development, static reservoir models are built by incorporating well logs, core, and seismic data. However, seismic amplitude data often cannot image small faults, resolve thin reservoirs, or discriminate subtle changes in reservoir properties. Seismic attributes aid in illuminating subtle faults and stratigraphic features but, analyzing numerous individual attributes can be tedious and may have limitations for revealing small petrophysical changes within a reservoir. Using the Maui 3D dataset acquired in the offshore Taranaki Basin of New Zealand, I generate seismic attributes that are sensitive to faults as well as attributes that are sensitive to reservoir properties. I use principal component analysis (PCA) and self-organizing maps (SOM) to integrate geological information from six geometric attributes into one classification volume, which shows small faults that could affect field compartmentalization. I also develop a machine learning workflow to combine the reservoir information from eight spectral instantaneous attributes into one clustered volume. This SOM classification shows the reservoir distribution and helps to understand reservoir quality and to illuminate thin baffles. During reservoir production, time-lapse (4D) seismic data can be acquired to monitor fluid movement and constrain simulation models. However, 4D seismic data suffer from seismic interference and tuning effects. Thus, it can be challenging to monitor fluid behavior within thin reservoirs and/or to illuminate baffles. As a result, the reservoir simulation model may not capture sufficient reservoir heterogeneity detail, which can lead to a mismatch between synthetic models and observed 4D seismic responses. Using the 4D seismic data from the Maui field, I develop a systematic workflow to carefully select the attributes that best show 4D changes and integrate the preproduction and postproduction reservoir information from multiple 4D spectral instantaneous attributes into 4D classification volumes. Changes in SOM cluster patterns between baseline and monitor surveys suggest production-related changes within good to poor quality reservoirs that are primarily caused by water saturation changes as the reservoir is being water swept. Geobodies derived from the SOM clusters are used to calculate recoverable gas reserves that are compared to production data.



Machine Learning, Reservoir Characterization, 4D Seismic