International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Data-Optimized Dry Rock Property Prediction Using Ensemble and Kernel-Based ML Methods

Authors :

Esmael Makarian¹ Hassanreza Ghasemitabar² Alireza Behinrad³ Mahdi Fathi⁴ Andisheh Alimoradi⁵ Ayub Elyasi⁶

1- Sahand (Tabriz) University of Technology 2- Shahrood University of Technology 3- Sahand (Tabriz) University of Technology 4- IKIU 5- IKIU 6- Department of Petroleum Engineering, College of Engineering, Knowledge University, Erbil 44001, Iraq

Keywords :

Dry Rock Module،Digital Rock Physics،Machine Learning،P–Wave Velocity

Abstract :

The accurate estimation of Dry bulk modulus (K) and Dry density (ρ) is of the utmost importance in digital rock physics (DRP); however, traditional approaches are data-demanding, time-consuming, and costly. This study strives to generate a novel insight to employ machine learning (ML) approaches to indirectly estimate K and ρ for a carbonate reservoir using P-wave velocity (VP) derived solely from well-log data. Two ensemble-based algorithms, including Random Forest (RF) and Gradient Boosting Regressor (GB), were employed alongside a kernel-based method represented by Support Vector Machine (SVM). The Dry density and Dry bulk modulus obtained from numerical calculations were subjected to a comprehensive multi-dimensional performance evaluation using five independent statistical metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), Median Absolute Error (MeAE), Explained Variance Score (EVS), and the coefficient of determination (R²). GB exhibited the highest performance, achieving validation R² values of 0.9211 for K and 0.9308 for ρ, along with the lowest mean absolute errors. Additionally, the analysis introduces a Composite Performance Index (CPI) to balance accuracy (R²) and stability (ΔMAE) across modulus and density targets, highlighting GB as the top performer (CPI ≈ 0.917), followed by RF (≈0.875) and SVM (≈0.785). These findings manifest that ensemble learning models, especially GB, capture the nonlinear relationship between VP and fundamental properties of rock. The method provides a fast, cost-effective, and input-optimized technique, making it particularly valuable in the DRP workflows where data availability is limited.