A seismic reservoir permeability prediction approach based on gaussian process machine learning.

Area of Science:

• Geophysics
• Petroleum Engineering
• Machine Learning

Background:

• Accurate permeability prediction from seismic attributes is crucial but challenging due to inherent uncertainties.
• Establishing a reliable inverse relationship between permeability and seismic attributes is key for reservoir characterization.
• Machine learning, particularly Gaussian Process (GP), offers potential but is limited by feature quantity and quality.

Purpose of the Study:

• To develop an integrated approach for enhanced permeability prediction using Gaussian Process (GP).
• To address challenges related to feature engineering and imbalanced datasets in permeability prediction.
• To improve the performance of GP models in predicting reservoir permeability.

Main Methods:

• Automatic generation of 222 extended features from three elastic attributes for GP training.
• Application of the synthetic minority oversampling technique (SMOTE) to handle imbalanced training samples.
• Utilizing Shapley values for feature importance assessment to select optimal predictors.

Main Results:

• The proposed integrated approach significantly enhances GP performance in permeability prediction.
• Feature importance analysis successfully identified key attributes for accurate prediction.
• Validation on a dolomite reservoir in Western China demonstrated the practical utility of the method.

Conclusions:

• The integrated GP approach, with automated feature engineering and selection, effectively improves permeability prediction accuracy.
• Addressing data imbalance and selecting relevant features are critical for successful machine learning applications in geophysics.
• This methodology provides a robust framework for reservoir characterization and management.