A Comparative Study for Machine-Learning-Based Methods for Log Prediction

To automate log interpretation at field scale, computational methods used to predict partially or entirely missing logs can be valuable. Such approaches could be potentially useful for correcting intervals of low-quality data, as well as predicting intervals outside the reservoir region where measured data are sparse. Furthermore, these capabilities may be useful as preconditioning for downstream analysis; e.g., to improve the class balance of different formation types during the log-seismic integration. The objective of this study is to compare the performance of several promising machine-learning (ML) methods for predicting missing logs. We include the following methods in the comparison: the window-based convolutional neural network autoencoder (WAE), the pointwise fully connected autoencoder (PAE), and the tree-based pointwise eXtreme Gradient Boosting (XGBoost). We also discuss the underlying assumptions and advantages of each method. With ML applications becoming increasingly popular among geoscientists, this study helps us understand the advantages and limitations of different regression methods used for log prediction and helps us better evaluate such ML approaches against traditional methods. For the various methods under consideration, we compare the computational complexity, model complexity (or model capacity), convergence rate during training, prediction error [root mean square error (RMSE) and mean absolute error (MAE)], and the analysis of both high- and low-scale feature reconstruction. We also note the unique aspects of each method. The study is conducted on well log data from multiple fields, and each field poses a variety of challenges, including the presence of coal, gas, radioactive sandstones, and highly rugose boreholes. We use the above-mentioned methods to predict one of the following logs: density, neutron porosity, gamma ray, or compressional slowness. We assume only one missing log at a given interval. Sequence-based methods can capture serial context information that allows for prediction in zones with depth-shifted logs, as well as using the information learned from nearby data to predict missing intervals. However, depending on the network capacity and the amount of training data, it can result in poor reconstruction of the short-scale features. On the other hand, while pointwise methods are less affected by nearby low-quality data, they may be challenged when the context information is important because of ambiguity in the logs.