Group Method of Data Handling (GMDH) Neural Network for Estimating Total Organic Carbon (TOC) and Hydrocarbon Potential Distribution (S1, S2) Using Well Logs

Recent advancement in computing capabilities has brought to light the application of machine learning methods in estimating geochemical data from well logs. The widely employed artificial neural network (ANN) has intrinsic problems in its application. Therefore, the objective of this study was to present a group method of data handling (GMDH) neural network as an improved alternative in predicting total organic carbon (TOC), S-1,S- and S-2 from well logs. The study used bulk density, sonic travel time, deep lateral resistivity log, gamma-ray, spontaneous potential, neutron porosity well logs as input variables to predict TOC, S-1, and S-2 of the Nondwa, Mbuo, and Mihambia Formations in the Triassic to mid-Jurassic of the Mandawa Basin in southeast Tanzania. The TOC prediction results indicated that the GMDH model trained well while generalizing better across the testing data than both ANN and Delta logR. Specifically, the GMDH provided TOC testing predictions having the least errors of 0.40 and 0.45 for mean square error (MSE) and mean absolute error (MAE), respectively, as compared to 1.27 and 0.81, 0.68 and 0.7, 1.4 and 0.89 obtained by backpropagation neural network (BPNN), radial basis function neural network (RBFNN), and Delta logR, respectively. For S-1 and S-2, the ANN models performed excellently during training but were unable to produce similar results when tested on the completely unseen well data. This represents a clear case of over-fitting by ANN. During testing, the GMDH avoided over-fitting and outperformed ANN by obtaining the least MSE of 0.04 and 1.16 and MAE of 0.07 for S-1 and S-2, respectively, while BPNN achieved MSE and MAE of 0.08 and 0.17 for S-1, 1.96, and 0.9 for S-2, and RBFNN obtained MSE and MAE of 0.15 and 0.25 for S-1 and 1.4 and 0.87 for S-2. Hence, the improved generalization performance of the GMDH makes it an improved form of a neural network for TOC, S-1,S- and S-2 prediction. The proposed model was further adopted to predict the geochemical data and determine the source rock quality for the East Lika-1 well, which has no core data.