DELINEATING SANDSTONE PROPERTY IN WEST TRYAL ROCKS USING ROCK PHYSICS GUIDED MACHINE LEARNING

• Main Author: Isma Aditya Saputra, Rino
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/69546
• Institution: Institut Teknologi Bandung

The sandstone in Mungaroo formation is believed to have different reservoir quality in each location of study area. It is important to map the sandstone property such as clay volume, porosity, and water saturation as 3D volume. The application of machine learning in geoscience that becomes popular to predict rock property is challenging to be applied, particularly using the dataset of study area. It is due to one of the main challenges of machine learning in geoscience such as supervised neural network is the requirement of labeled data input for the training purpose to determine statistical operator. The more data input available for the training purpose, the more reliable statistical operator used for best quality of properties prediction. To tackle these issues, this research discusses a form of hybrid theory-guided data science approach or rock physics-guided machine learning. The theory mentioned in this approach is based on rock physics model theory and use it as input for the data science/machine learning part. There are 2 wells available in this research; the WTR-4A well will be used for simulation and training purpose, while the second well WTR-2 will be used as a blind well. WTR-4A well is used to simulate many pseudo-wells based on statistics in the project area. The petrophysical properties, such as thickness, saturation, and porosity are all varied to create a well-sampled data set. Based on rock physics mode theory, the elastic pseudo-logs and synthetic seismic data are generated from all variations of petrophysical properties based on realistic geological conditions. The pseudo-well logs and synthetic data from that simulation are then used as input for neural network training. The operator derived from that pseudo-well training is then applied to the real seismic data to predict rock property volumes such as clay volume, porosity, and water saturation. The results from rock physics-guided machine learning firstly generate the elastic properties (acoustic impedance and Vp/Vs) and it will be compared to the common/well-established method for seismic reservoir characterization (seismic simultaneous inversion). Based on the result comparison from those two methods, the rock physics guided machine learning results look more promising as it has v better continuity on sandstone reservoir layer, good correlation with the blind well, and more detail (better separation between sand and shale layers). The combination of both elastic properties (acoustic impedance and Vp/Vs) can help the interpretation to distinguish the main sandstone reservoirs that contain gas (M, N, and O sand) with brine-saturated sandstone (T-sand). Beside predicting the elastic properties, another 3D volumes that is generated are petrophysical properties (volume of clay, porosity, and water saturation). These petrophysical volumes are also validated with the blind well and it has good correlation and the characteristic trend aligned with the elastic properties. The maps of each petrophysical properties are generated to analyze the lateral characters for M, N, and O sandstone layers. For the O layer, it has relatively low Vcl for the whole area while the high porosity and low water saturation mostly located on the southern part, especially south-west. In the meantime, the N layer has good quality reservoir on the south-eastern part while the northern area that has low Vcl indicates the low porosity and high water saturation. The most upper part of reservoir (M-sand), all the three petrpophysical properties have the same trend and all the good quality reservoir located on the southern part with southwest-northeast trend.