Hybrid Traditional Interpretation and Machine Learning-Based Log Reconstruction for Enhanced Petrophysical Evaluation of Hartha Formation, of XS Oil Field, Central Iraq

Abstract

To improve reservoir characterization of the Hartha Formation in the XS oil field in Central Iraq, this work presents a hybrid technique that combines conventional petrophysical interpretation with machine-learning-based log reconstruction. The consistency and reliability of the input dataset were enhanced by reconstructing missing well-log data using a random forest regression model in conjunction with a shift-and-match depth-alignment approach. The model was trained and validated using the reference well EB-4. The trained process was then applied to well EB-1, EB-5, EB-6, and EB-54, which had incomplete critical log types such GR, RHOB, NPHI, and CALI. The capacity of a Random Forest model to capture nonlinear interactions among petrophysical parameters in a heterogeneous carbonate system was confirmed by its excellent prediction accuracy, R² 0.78-0.90. Shale volume, water, and hydrocarbon saturation, and net pay thickness were calculated using Schlumberger Techlog after the logs were rebuilt. The results showed that while water-bearing intervals dominate Ha2 and Ha3-d, the Ha3-d subunit, especially Ha3-a and Ha3-c, has the best reservoir characteristics with high porosity >0.25 v/v and hydrocarbon saturation (>0.45 v/v). The work shows that combining classical interpretation with AI-driven prediction fills in data gaps, lowers uncertainty, and quantitative understanding of reservoir heterogeneity.