Similarity Learning for Well Logs Prediction Using Machine Learning Algorithms

Abstract

Determining and predicting reservoir formation properties for newly drilled wells represents a significant challenge for oil and gas companies. Extensive well logs are available only while or after drilling, and thus they bear substantial financial, technical, and operational risks. We propose a new machine learning data-based model for determining well properties similarity and further derive and predict well logs before drilling in a specific geological context. Our model starts with selecting crucial well intervals and aggregation of vital features that determine the petrophysical properties related to particular well layers. Then, a machine-learning algorithm uses this info as input to provide a similarity score between wells. Our fast-to-train nonlinear data-based model is a variant of gradient boosting. We show that this approach can work well in complex scenarios with missing data and inconsistent similarity measures. We compare the modern machine learning algorithms for the evaluation of well similarity models based on aggregated features. The algorithms include gradient boosting and baseline logistic regression models. Our assessment for a real well log dataset via group cross-validation demonstrates that the gradient boosting model pretty accurately identifies well similarity. The receiver operating characteristic quality metric (ROC AUC) is 0.824. The developed similarity learning framework provides a data-driven approach towards estimating well logs for planned and newly drilled wells. Therefore, it allows prediction, improves determination, and can drive an optimal selection of log measurements to be executed in a new well in a specific field / geological context.