AI Based Prediction Of Relative Permeability Curves From Well Logs And Core Data For Enhanced Reservoir Simulation Accuray

Abstract

This dissertation presents an AI based methodology to predict relative permeability Corey parameters from well logs and core data to improve reservoir simulation accuracy for the Hassi Berkine (Block 404) field. After reviewing reservoir characterization, SCAL techniques, and relative permeability theory, the study builds a dataset of 100 samples combining petrophysical logs (porosity, permeability, gamma ray, density, neutron porosity, resistivity, shale volume, FZI) and core derived targets (Swc, Sor, krw_end, kro_end, nw, no). Data preprocessing (missing value removal and 0–1 scaling), feature selection (correlation analysis and importance ranking), and model comparison form the workflow. Four algorithms Linear Regression, ANN, Random Forest, and XGBoost are trained and evaluated using RMSE and R² metrics and curve alignment plots. Linear regression failed to capture nonlinearity (Total R² ≈ 0.044), while ANN and unconstrained Random Forest achieved near perfect fits on training data (R² ≈ 0.999–1.000) but risked overfitting given the small sample size. XGBoost provided the best balance between accuracy and generalization (high R², low RMSE). The work demonstrates that machine learning can generate continuous, field wide relative permeability curves from logs, reducing reliance on sparse SCAL data; however, robust validation, dataset expansion, and hyperparameter tuning are required before operational deployment in Petrel based reservoir models.