No abstract

No abstract

Log interpretation is the task of analyzing and processing well logs to generate the subsurface properties around wells. A direct application of machine learning (ML) to this task is to train an ML model for predicting properties in target wells given well logs (data) and properties (labels) in a set of training wells in the same field and/or region. Our ML model of choice for predicting the desired properties is the decision tree-based learning algorithm called random forests (RF). We also devise a mechanism to automatically tune the hyperparameters of this algorithm depending on the data in the training wells. This eliminates the tedious task of carefully tuning the hyperparameters for every new set of training wells and provides a one-click solution. In addition to predicting the properties, we compute the uncertainty in the predicted properties in the form of prediction intervals using the concept of quantile regression forests (QRF). We test our workflow on two use cases. First, we consider a petrophysics use case on an unconventional land dataset to predict the petrophysical properties such as water saturation, total porosity, volume of clay, and total organic carbon from petrophysics logs. Then, we consider a geomechanics use case on a conventional offshore dataset to predict the lithology, pore pressure, and rock mechanical properties. We obtain a good prediction performance on both use cases. The uncertainty estimates also complement the ML model's prediction of the properties by explaining the various correlations that are found to be existing among them based on domain knowledge. The entire workflow of automating the tuning of hyperparameters and training the ML model to predict the properties along with its estimate of uncertainty provide a complete solution to apply the ML workflow for automated log interpretation.