This study explores the capability of a Multilayer Perceptron (MLP) neural network (NN)machine to predict missing or expensive core rocks and well-log data measurements such as the Total Organic Carbon (TOC) and Elemental Capture Spectroscopy (ECS) measurements. Data of boreholes drilled in the Lower Barnett Shale and Bakken oil and gas fields, located in the USA, are used. TOC estimation is first addressed using the Schmoker method in the Barnett Shale gas and Bakken oil reservoirs, followed by the implementation of MLP NNs trained with various learning algorithms such as the Hidden Weight Optimisation, the Conjugate Gradient, and the Levenberg-Marquardt. Input data include standard well logs such as sonic, gamma ray, resistivity, and neutron porosity. The MLP models are validated and generalised using both horizontal and vertical well data. Furthermore, ECS data prediction is performed using MLPs trained on elementary analysis-derived log parameters, offering a cost-effective alternative to direct ECS logging. The results demonstrate that the efficiency and reliability of MLP-based approaches in enhancing geochemical and petrophysical characterisation of subsurface formations is conditioned by the choice of the learning algorithm, the reservoir complexity, number of wells, and their distribution.
On the capability of multilayer perceptron to predict total organic carbon and elemental capture spectroscopy data in unconventional hydrocarbon reservoirs: the case of the Barnett Shale and Bakken oil field
Abstract: