An Improved Computational Learning-Based Model for Estimating Total Organic Carbon in Unconventional Shale Reservoirs

Abstract

Unconventional resources have emerged as one of the crucial alternatives to the rapidly depleting of conventional hydrocarbon resources. The hydrocarbon potential of shale source rocks is assessed by the percentage of the organic index such as total organic carbon (TOC). Correct estimation of TOC is very important since minor deviations in anticipated results can lead to wastage of investments and time. A slight improvement in estimation practices, on the other hand, can increase the value of an exploration project. Therefore, the objective of this study is to present an improved classification and regression tree (CART) computational learning-based model as an improved alternative in estimating TOC from well logging data. Conventional well logs suite of bulk density, gamma-ray, deep resistivity, sonic transit time, spontaneous potential, and neutron porosity from Mihambia, Mbuo and, Nondwa, Formations of the Mandawa Basin Tanzania, were used as input variables. Results from the developed CART TOC model were compared with the random forest (RF) and backpropagation neural network (BPNN). It was observed that the proposed CART model trained better while generalizing better through unused testing data compared with RF and BPNN. CART model achieved R, RMSE, and MAPE values of 0.9615, 0.0840, and 0.5035 for training and 0.9703, 0.1162, and 0.3722 for testing respectively. The proposed model work with higher accuracy with the sensitivity analysis indicating that gamma-ray, deep resistivity, and sonic transit time significantly influenced the model outcome.