DEVELOPMENT OF TREE-BASED ENSEMBLE LEARNING ALGORITHMS FOR ESTIMATING TOTAL ORGANIC CARBON FROM WIRELINE LOGS

To evaluate the hydrocarbon generation potential of a rock strata, total organic carbon (TOC) is a significant factor. TOC estimation is considered as a challenge from well logs and direct measurement in laboratory from rock specimens is costly and time-consuming. Therefore, due to the complex and nonlinear relationship between well logs and TOC, researchers have begun to use artificial intelligence (AI) techniques. Prediction from Passey method is low and AI techniques such as Artificial Neural Network (ANN), Support Vector Machine (SVM) gets trapped in local optima resulting in overfitting and heavy computational work and even error if the technique isn’t reasonable. In this thesis work, for the TOC prediction we proposed seven AI algorithm - Artificial Neural Netwrok (ANN), four efficient tree-based ensemble techniques that includes Random Forest (RF), Extra Trees (extremely randomized trees) (ET), Gradient Boosting (GB) and Extremely Gradient Boosting (XGB) and two hybrid AI models that includes Genetic algorithm- Artificial Neural Network (GA-ANN) and Particle Swarm Optimization- Artificial Neural Network (PSO-ANN) have been used. Among seven algorithms studied in this work, the four tree-based ensemble models are capable of fitting highly non-linear data and requires minimum data pre-processing. Specifically, 205 data points and seven well logs namely GR, DT, RHOB, SP, NPHI, LLD, and LLS were used from the Goldwyer Formation of the Canning Basin for training and testing the seven AI models to evaluate their efficiency and provide comparable results during the TOC estimation. From results it is validated that the accuracy of the tree-based ensemble techniques is at exemplary level for the TOC content estimation where the XGB model for training and testing data sets outperformed all the other AI models especially all other tree-based ensemble techniques i.e., RF, ET and GB. These robust tree-based ensemble models not only protect overfitting but has achieved better prediction results while dealing with the multidimensional data. Finally, some possible combinations are proposed that have not yet been investigated.