Behnam Jafarpour

University of Southern California

Integrating data and physics to enhance production prediction in unconventional reservoirs

The complexity of fluid flow physics in hydraulically fractured unconventional reservoirs limits the ability of models in predicting the observed field performance data. Prediction errors are caused by a combination of incorrect description of the physics, inability to include relevant field measurements (well logs and drilling parameters) that contain important information about reservoir condition and reservoir characteristics, and errors in specifying model inputs. Data-driven models are powerful tools to develop predictive models, especially when physics-based models are not available or reliable. I will present neural network (NN) architecture as a data-driven tool to learn the errors/residuals in physics-based prediction of production performance in unconventional reservoirs. Using a training dataset consisting of the discrepancy between physics predicted and observed production data, together with formation and completions parameters for each well, the NN architecture learns the prediction residuals/errors of the physics model. The method leverages data-driven models to account for prediction errors originating from potentially missing physical properties and processes, as well as simulation input parameters. Examples from Bakken Play in North Dakota are used to illustrate the method.