Real-Time Machine Learning Application for Formation Tops and Lithology Prediction

Abstract

During the drilling operation, the drill string is subjected to different geological formations which have distinct lithological characteristics that greatly affect the drilling performance and may ultimately result in increased costs of the project. The lithology of a formation can vary significantly, thus it is of paramount importance to accurately detect lithology changes and formation tops while drilling. In order to do so, geologic data and logs are often utilized by experts and operators to identify lithological variations. Machine learning algorithms and random forest have been employed in recent years to improve the process of lithology prediction, enabling more accurate results at faster rates. Machine learning-based systems incorporate a wide range of indicators such as rock types, mineral composition, sedimentary structures and microfossils for efficient lithology prediction. Additionally, random forest classifiers are beneficial due to their robustness with respect to outliers as well as their ability to capture complex relationships between variables from multivariate input datasets. With this approach, an effective operational strategy can be formulated based on the identified formation lithology in order to reduce incident costs associated with unexpected wellbore issues or instability caused by lithological changes. This technique also provides valuable insight into understanding subsurface conditions for more efficient resource exploration and production operations. limitations and drawbacks of this approach as cost and lag time. The current study proposed an intelligent machine learning solution for auto-detecting drilled formation tops and lithology types while drilling in real-time utilizing drilling surface data. Machine learning techniques are technically employed for developing real-time prediction models for the formation tops and lithology type from the surface drilling data as weight on bit, drill string speed, torque, pumping pressure and rate, and drilling penetration rate. This study implemented random forest and decision trees as two machine learning classifiers to develop real-time models using a data set of composite lithology schemes of five drilled formations. The methodology approach presents a comprehensive layout for data collection, preprocessing, data statistics and analytics, feature engineering, model development, parameters optimization, and prediction performance evaluation. The results showed a high prediction performance for the models for training and testing with overall accuracy higher than 95 through detecting complex lithology schemes. Predicting the drilled formation's tops and lithology while drilling in real-time through the developed solution will provide a technical guide for optimizing the drilling parameters for better drilling performance and optimized mechanical-specific energy to have a safe operation and cost savings.