Formation permeability estimation using mud loss data by deep learning.

Area of Science:

• Petroleum Engineering
• Machine Learning
• Geoscience

Background:

• Permeability estimation is crucial for reservoir assessment and hydrocarbon extraction.
• Existing methods for permeability evaluation may be inaccurate or unavailable.
• Mud loss data, often overlooked, presents a potential source for permeability estimation.

Purpose of the Study:

• To develop and validate a novel method for estimating formation permeability using mud loss data.
• To apply deep learning techniques for accurate permeability prediction.
• To explore the utility of real-time drilling data for reservoir characterization.

Main Methods:

• Generated mud loss rate data using a reservoir simulator with varying reservoir and drilling parameters.
• Employed one-dimensional convolutional neural networks (1D-CNN) for permeability estimation.
• Utilized a novel Deep Jointly Informed Neural Network (DJINN) model, integrating neural networks and decision trees.

Main Results:

• 1D-CNN achieved high accuracy in permeability estimation (R²=0.970 training, R²=0.964 testing).
• The DJINN model outperformed 1D-CNN, demonstrating superior accuracy (R²=0.978 training, R²=0.972 testing).
• Validated the correlation coefficients of generated data to ensure reliability under real-world conditions.

Conclusions:

• Mud loss data can be effectively utilized with deep learning to accurately estimate formation permeability.
• The DJINN model offers a more accurate approach compared to 1D-CNN for this application.
• This methodology provides new applications for drilling data, enabling petroleum engineers to improve reservoir design and characterization.