Comprehensive input models and machine learning methods to improve permeability prediction.

Area of Science:

• Geosciences
• Machine Learning
• Petrophysics

Background:

• Permeability estimation is crucial for reservoir characterization.
• Traditional methods often face limitations in accuracy and efficiency.
• Machine learning offers a promising alternative for improved permeability prediction.

Purpose of the Study:

• To evaluate machine learning techniques for permeability estimation.
• To determine the optimal selection of input well logs (gamma ray, resistivity, porosity, density, sonic, neutron porosity).
• To compare the performance of five machine learning algorithms: Extreme Learning Machine (ELM), Random Forest (RF), Gradient Boosting (GB), K-Nearest Neighbor (KNN), and Multilayer Perceptron (MLP).

Main Methods:

• Developed 57 unique models using combinations of six input well logs.
• Tested models using five distinct machine learning algorithms.
• Generated 285 unique permeability predictions.
• Validated model performance on a blind well dataset.

Main Results:

• Random Forest (RF) achieved the highest correlation coefficient (0.925) but a higher average error (0.196).
• Extreme Learning Machine (ELM) demonstrated the lowest average error (0.083) with a correlation of 0.871.
• Gradient Boosting (GB) and RF models proved highly effective in blind well testing, yielding R² values of 0.92 and 0.90, respectively.
• A precision-correlation trade-off was observed among different models.

Conclusions:

• The selection of appropriate machine learning algorithms and input data is critical for accurate permeability estimation.
• Gradient Boosting and Random Forest are highly effective for permeability prediction, even in untested geological scenarios.
• Machine learning provides a robust framework for enhancing the reliability and accuracy of permeability models in geosciences.