Predicting drilling fluid lost circulation volume and radius using petrophysical well logs and machine learning...

Area of Science:

• Petroleum Engineering
• Machine Learning Applications
• Drilling Operations Optimization

Background:

• Lost circulation is a major challenge in drilling, causing significant non-productive time and financial losses.
• Accurate prediction of loss zones, volumes, and fluid invasion is crucial for effective management.
• Existing methods often lack the precision needed for proactive mitigation.

Purpose of the Study:

• To develop and evaluate an integrated machine learning framework for predicting lost circulation.
• To compare the performance of multi-layer perceptron (MLP), Kolmogorov-Arnold network (KAN), and a hybrid MLP-KAN model.
• To identify key predictive features influencing mud invasion and loss zones.

Main Methods:

• Integration of petrophysical well logs, drilling operational parameters, and mud characteristics.
• Development of three machine learning models: MLP, KAN, and a hybrid MLP-KAN architecture.
• Utilized a dataset of 246,620 data points from 73 wells in an Iranian oil field.

Main Results:

• The hybrid MLP-KAN model achieved superior performance with R²=0.9445 and RMSE=0.0143, outperforming standalone MLP (R²=0.9054) and KAN (R²=0.9075).
• Successfully identified three primary mud invasion zones at specific depths (1800-2000m, 2300-2600m, 2800-3000m) and quantified invasion volumes/radii.
• SHAP analysis highlighted sonic transit time (DT), bit size, mud weight, and calcium content as significant predictive features.

Conclusions:

• The hybrid machine learning framework offers a robust and interpretable tool for proactive lost circulation management.
• Enables zone-specific prevention strategies and optimization of drilling fluid programs, reducing risks and costs.
• Provides a scalable solution for enhancing drilling efficiency by bridging data analytics and practical field applications.