A Self-Adaptive Artificial Neural Network Technique to Predict Total Organic Carbon (TOC) Based on Well Logs

Abstract

Determination of total organic carbon (TOC) is a key method of characterizing shale reservoirs. The conventional method for TOC determination using cores from shale reservoirs is time-consuming and costly. TOC can be estimated by an indirect method using petrophysical well logs. The existing models assume a linear relationship between the well logs and TOC and have a high error and low correlation coefficients (CC) between the actual and predicted TOC. The first goal of this study is to apply a self-adaptive differential evolution (SaDE) optimizing method to determine the best combination of artificial neural network (ANN) parameters (number of hidden layers, number of neurons in each layer, training function, transferring function, and the training over testing ratio). The second goal is to develop a new empirical correlation that can be used to estimate TOC using well logs based on the optimized SaDE-ANN model. Four-hundred and sixty data points from Barnett shale were used for training and testing the developed SaDE-ANN model. Another set of data (29 data points) of Duvernay shale was used to compare the developed TOC correlation with the previous models. The obtained results show that the developed SaDE-ANN model predicted the TOC using only well logs: gamma ray (GR), compressional time (DT), deep resistivity (RD), and bulk density (RHOB) with a high accuracy (a CC of 0.99 and the average absolute percentage error (AAPE) of 6%). The developed TOC correlation outperformed the models proposed by Wang et al. (Mar Pet Geol 70:304–319, 2016. https://doi.org/10.1016/j.marpetgeo.2015.11.023 ) and Abdulhamid et al. (Int J Coal Geol 179(15):72–80, 2017). The new empirical correlation for TOC estimation reduced AAPE by 67% as compared with the ANN model developed by Abdulhamid et al. (2017) for the Duvernay formation. The developed TOC correlation is simple and can be applied using any computer without the need for the ANN model or special software. The developed technique will help reservoir engineers and geologists to estimate the TOC values using only the well logs with a high accuracy.