Abstract
This paper discusses successful implementation of machine learning techniques like Neural networks for prediction of reservoir facies in a highly variable clastic depositional area with limited core & log data. These predicted facies were then populated into the static model for robust field development and future well placement in Beta field of Middle Indus Basin, Pakistan.
In the study area, only 1 well outside the field had reservoir core data. This data and poro-perm facies were combined with log data to train the neural network (1D convolutional neural net). After getting an excellent match at known points, trained network was used to predict facies at Beta-1 well (exploratory well that discovered the field) utilizing the available well logs. Predicted facies were then distributed statistically using variogram (based on geological and depositional architecture of the area) and co-kriged with seismic amplitude distribution within the Beta field area.
The predicted facies were used as one of the inputs into the Reservoir Static model. Post modelling, favorable undrained area was identified in the Beta field and next development well was drilled. Well results showed that predicted facies matched with new well results by more than 70%. The results clearly demonstrate the effectiveness of machine learning techniques in predicting reservoir facies. This successful demonstration of facies prediction highlights the significance and power of implementation of machine learning methods when used with geological & geophysical data. This approach has allowed for better understanding of field dynamics, aided in better well placement, and ultimately improved production performance in the Beta field.
