Prior knowledge of reservoir fluid type and properties aids in selecting and optimizing completion and surface facilities. Fluid properties prediction has an impact on in-place volumes and reservoir performance management including optimized well placement. We present a data-driven fluid variation modeling approach using machine learning. The aim is to predict the fluid type and oil API gravity for a given location and depth and optimize the completion design for the Eagle Ford shale.
Data from 9400 Eagle Ford shale wells were compiled, cleaned, and analyzed. Data was then divided into training and test sets. The test set was set aside for validation to prevent any training bias. Data visualization and statistical analysis was carried out, which revealed patterns and features within the training data. Three separate artificial neural networks (ANNs) were then constructed on those features, and a supervised learning algorithm was employed to train on the training set.
The first ANN predicts the oil API gravity based on a given coordinate: latitude, longitude and depth information. This network uses Mean Squared Error (MSE) loss function with the Root Mean Squared (RMS) regression optimizer. ANN-1 reported an error of 2.4 API which is well within process dependency of the API measurements and within the potential experimental errors. The second ANN predicts the most likely fluid type along with the probability, which can be used as a measure of confidence. ANN-2 uses the categorical cross-entropy loss function with the Adam optimizer (Kingma (2014)). Finally, ANN-3 predicts the hydrocarbon production of the first 12 months based on the well location, lateral length, depth, number of stages, proppant volume and gel volume. All three models were then validated on the test set, and a good match was obtained. Based on the data-driven models, an optimization scheme was created to maximize cash flow from the first 12 months of production based on varying the lateral length, the number of stages, proppant volume, and gel volume used. The resulting optimum parameters are then represented visually on the map of Eagle Ford, along with oil and gas production, and cash flow.
Even though the presented method was trained for Eagle Ford, data from other formations can be incorporated and re-trained, including other proxies for every additional basin, to create a general neural network predictive model on all formations; or to create smaller networks that would make accurate predictions within the specified formation. This approach will lead to a continuously improving and learning process for each additional field and play.