Application of artificial intelligence techniques to predict log at gaps: a case study

The data driven approach to predict specific log/s from available data is essential when there are log gaps due to operational issues. The imputation at missing interval is crucial for rock physics modeling or synthetic seismic tie, as those studies require complete log. The artificial intelligence techniques such as random forest (RF) or artificial neural network (ANN) are effective to predict log from available log data. These algorithms have the advantages like scalability, liner/non-linear correlation between input logs, less chance of model overfitting, robust architecture, efficiency to reduce loss function etc. The selection of the best algorithm for specific area requires multiple testing of the predicted values. In this study, random forest and artificial neural network algorithms are used to predict sonic logs at missing intervals of approximately 20 m. As the missing interval covers reservoir zone, it poses limitation to carry out reservoir characterization and geomechanical modeling at the interval. The dataset for training and testing intervals was selected based on the quality of the data and geological input. The optimized hyper parameters were listed for both algorithms. To capture the uncertainties associated with the predicted sonic log from random forest, quantile percentage error was calculated. The residual errors like Mean Average Percentage (MAPE) and Root Mean Square Error (RMSE) were calculated on training data. The MAPE and RMSE errors are within an acceptable range, which are 1-2.5 and 0.01 respectively. The predicted sonic logs were plotted in velocity-porosity space with existing rockphysics model lines. It showed the value range of 2000-2300 m/s. The plot depicts that the ANN predicted values fall better on friable sand line compared to RF. Also, in the synthetic-seismic tie, the ANN output shows better correlation (similar to 70%) than with RF output (similar to 63%). The study shows a pragmatic approach to apply artificial intelligence techniques to train available log data for prediction of sonic log and further applications in rock physics model, geomechanical model and synthetic seismic correlation.