Deep learning integrated approach for hydrocarbon source rock evaluation and geochemical indicators prediction in the Jurassic - Paleogene of the Mandawa basin, SE Tanzania

Abstract

The world's energy demands are growing at an unprecedented rate, and the exploration of new hydrocarbon sources is more important than ever. Therefore, the objective of this study was first to quantitatively analyze hydrocarbon source rock potentiality of the Triassic-Jurassic of Mandawa Basin based on the generalized group method of data handling neural network (g-GMDH), Machine learning, and Geochemical using well logs data. Then a novel g-GMDH was presented to predict a continuous geochemical log profile of TOC, Tmax, S1, and S2. It was observed that the basin's hydrocarbon source rocks are classified as fair to very good source rocks with TOC contents ranging from 0.5 to 8.7 wt%. The source rocks contain mixed kerogen type II and III, which are oil and gas-prone, ranging from immature to mature source rocks. The results of the predictive models indicated that the g-GMDH model trained better whilst generalizing well throughout the testing data than both GPR and SVM models. Specifically, the g-GMDH when tested on unseen data had the least value of MSE = 0.18, 2.35, 0.08, and 61.74 for TOC, Tmax, S1, and S2 respectively, and MAE = 0.45, 1.37, 0.17 and 11.55 for TOC, Tmax, S1 and S2 respectively. The g-GMDH model was further applied to assess the source rock and predict the geochemical information in the East Lika well, which lacks core data. The proposed model can offer rapid and real-time values of geochemical indicators and are independent of laboratory-dependent parameters therefore, can be adopted as an improved technique for evaluating source rocks in frontier basins.