Journal of Engineering Research and Reports

The rate of fluid flow through reservoir rocks is determined by permeability, one of the key characteristics of a reservoir. Production forecasts, history matching, and robust reservoir simulation all depend heavily on accurate permeability estimates. It can be difficult to build trustworthy permeability models because of the inherent variety of permeability at various sizes and the scarcity of core data. To overcome these obstacles, this work uses a variety of machine learning techniques, including Support Vector Regression (SVR), Random Forest (RF), XGBoost, and LightGBM, to predict lab-measured core permeability from frequently obtained well logs. A dataset that represented the clastic formations (X- Field) was considered. The resilience of this technique under various geological settings could be assessed using the X-field dataset, which included three wells spread across three distinct reservoirs. This approach relies heavily on feature engineering, especially when it comes to integrating vertical variability. Taking into account the smoothing effect of well logs over small-scale heterogeneities and the significance of spatial context, measurements from nearby well log readings were added to the models. By taking into consideration nearby depositional environments and shared geological history, this increased prediction accuracy. The Bara, Kue, and Yorkiri formations had blind test R2 ratings of up to 0.84, 0.76, and 0.78 for the X- Field, respectively. Even if these results are satisfactory, they show how machine-learning techniques can be used to accurately estimate permeability and emphasize the need for feature engineering. This work argues that although automated feature engineering using machine learning shows potential, human intervention—more especially, the incorporation of geographical context—can still greatly improve predictions. It may be the goal of future developments to incorporate this spatial awareness into machine learning algorithms.