Document Type : Original Research Paper
Authors
1 University of Benghazi , Department of earth science , Benghazi , Libya .
2 Department Petroleum Engineering, Faculty of Engineering, Bright Star University, Elbrega, Libya.
3 Arabian Gulf Oil Company (AGOCO)
Abstract
Permeability estimation is an essential phase in assessing the hydrocarbon potential within porous media and designing reservoir management methods. Recently, machine learning (ML) methodologies have gained prominence in the prediction of permeability. The initial stage in constructing highly reliable ML models is to identify the optimum combinations of input logs, as permeability is a highly sensitive parameter; this step is essential and can influence model accuracy. While feature engineering methods provide valuable insights in selecting suitable input logs, the effectiveness of these logs or their combinations remains underexplored, particularly in the context of high-heterogeneity reservoirs. The current study intends to save time by evaluating the effectiveness of twelve distinct models, each constructed using a Multi-Layer Perceptron (MLP), based on various combinations of input logs using data from the Nubian reservoir, Sirt Basin, Libya. The methodology involved several steps, including preprocessing, splitting, optimization, and validation. The findings demonstrate that single-input logs, mainly the Gamma-ray (GR), bulk density (RHOB), and sonic logs (DT), exhibited higher correlation coefficients compared to the multiple log combinations. The GR model attained the best R² of 0.994, indicating its sensitivity in capturing non-linear relationships. On the other hand, multi-log models achieved variable accuracy, resulting in increased learning complexity. The study highlights the efficiency of selecting the optimal combination of input logs, providing practical guidance for ML-based permeability prediction in heterogeneous reservoirs.
Keywords
Main Subjects
)