Unconventional reservoirs are continuously facing severe technical challenges in safe drilling, successful completion, and effective fracturing to unlock their potential. One of the key gaps is accurately estimating the geomechanical properties (Young's modulus, Poisson ratio, etc.) due to its complex structure, reservoir heterogeneity, and insufficient borehole information. While these properties could be calculated from sonic logs, it often results in log deficiency and a high recovery cost of incomplete datasets. To fill the gap accurately, we propose both classical machine learning and deep neural network to estimate and predict the geomechanical properties of the Permian Basin.

The log-derived prediction algorithm includes (a) Single-Well prediction, 75% of log data of a single well is used as a specimen for training the Bi-LSTM, and the rest 25% of data of the same well is used for testing, and (b) Cross well prediction, a group of wells from the same region is divided into training and testing. The logs used in this work were collected from seven Permian Basin wells gamma-ray, bulk density, resistivity, etc. Finally, we employed four various machine learning (ML) algorithms (Decision Tree, Ada Boost, kNN, and Random Forest) to compare and investigate the efficacy of the deep neural network in predicting geomechanical properties.

The results demonstrate promising predictions of the geomechanical properties for the Permian Basin using ML and deep neural networks. The highest performance for a single well prediction using the ML models with an R2 value of 99.90%. In a deep neural network, Bi-LSTM performs superior with an accuracy of 92.5%. The highest average accuracy obtained for single well prediction is 90.7%. Cross well prediction performed superior for all wells compared to the single well prediction and for both known and completely unknown data sets. While datasets are incomplete, these results demonstrate the excellent ability of machine learning models trained on adequate data sets to generate a realistic prediction.

Given adequate training data, ML-based models will likely be able to fill the information gap by making an accurate prediction. Adoption of machine learning models for predicting geomechanical properties could have extensive impacts on both treatment design and the role of the geomechanicist. This research demonstrates the application of the Bi-LSTM model for the Permian Basin by accurately predicting geomechanical properties which can be used for both conventional and unconventional reservoirs to reduce cost and a considerable amount of time in completion workflows and thus increase the hydrocarbon recovery. Operators could leverage this promising deep learning model utilizing it as an automated handy tool to audit fracture interpretations quickly where datasets are incomplete.

This content is only available via PDF.
You can access this article if you purchase or spend a download.