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

Improving data quality during log preprocessing is an important task that can consume most of the time of the petrophysicist, with a high impact on the final interpretation. As part of the initiative to increase automation and homogeneity in the data completeness of logs in a field, we organized a systematic comparison of multiple regression models that provided successful predictions of wellbore logs. These approaches can be potentially valuable when extrapolating measurements available on a few wells to a more extensive set of wellbores, predicting low-quality data intervals, and increasing the availability of complete data sets. This study aims to compare the performance of three promising machine-learning (ML) methods when predicting one of the following curves: density, neutron porosity, and compressional slowness curves. We view the need to evaluate models that could provide answers even in the presence of multiple missing logs or logs with alteration, which is a common scenario in petrophysics. Because of that, we built a comparison based on three ML methods that can handle those issues: window-based convolutional neural network autoencoder (WAE), pointwise fully connected autoencoder (PAE), and eXtreme Gradient Boosting (XGBoost). We developed the PAE and WAE methods to handle challenging scenarios of interest, and we used the original implementation of XGBoost, which is already built to handle missing values. We compare the computational complexity, prediction errors [root mean square error (RMSE) and mean absolute error (MAE)], Pearson’s correlation, peak signal-to-noise ratio (PSNR), and the visual analysis of both high- and low-scale feature reconstruction, conducting the comparison in two field data sets. We also used the same methods to predict photoelectric factors and interpreted formation properties such as total organic content in multiple field data sets.