Machine-Learning Model for the Prediction of Lithology Porosity from Surface Drilling Parameters

Abstract

AbstractIntroductionThe accurate characterization of the lithology porosity is critical for geological interpretation and decision making in petroleum exploration. For this, wireline logging (including sonic, neutron porosity, and density, among other logs) is often used for the characterization of geophysical data performed as a function of wellbore depth. The common practice in the oil and gas industry is to perform the wireline logging for every new well, which is a lengthy and expensive operation. Therefore, the objective of this study is to use the historical logging data and surface drilling parameters to derive machine-learning (ML) models able to identify the different lithology classifications.MethodologyWe used historical logging data and surface drilling parameters to derive ML models to predict the following lithology classification: 1) porous gas, 2) porous wet, 3) tight sand, and 4) shaly sand. These models can predict these classifications without running wireline logs in the new wells. In this approach, the four lithology classifications are defined from the sonic, neutron porosity, gamma-ray, and density logs from historical data and are considered as the learning target/labels for the ML model. Therefore, the ML model learns the relationship between the surface drilling parameters and mud weight with their respective lithology classification. Finally, the model is capable of being executed in real-time, improving crew decision making.ResultsThe results obtained from a stratified 5-fold cross-validation technique demonstrated that the random forest model was able to learn from the data with an accurate classification for the four lithology porosity categories. The derived ML model obtained an average of 89.66% and 89.20% for precision and recall, respectively.NoveltyAlthough many studies have suggested the use of ML to imputing logging data, the inputs of these models are the data from other logs. Conversely, our proposed approach utilizes the wireline logging data only during the training of the model for assigning the porosity classification as labels. As such, the model learns the relationship between drilling parameters and the associated labels. This approach not only simplifies the learning of the ML but eliminates the need to run wireline logging in new wells, considerably reducing time and costs.