Deep Learning Enhances Transfer Learning for Tight-Oil Production Prediction

Abstract

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper URTeC 2021-5688, “Physics-Assisted Transfer Learning for Production Prediction in Unconventional Reservoirs,” by J. Cornelio, SPE, S. Mohd Razak, SPE, and A. Jahandideh, University of Southern California, et al. The paper has not been peer reviewed. _ A physics-assisted deep-learning model is presented to facilitate transfer learning in unconventional reservoirs by integrating the complementary strengths of physics-based and data-driven predictive models. The developed model uses a deep-learning architecture to map formation properties to their corresponding production responses using a low-dimensional feature space representation. The results indicate that physics-based simulated data can facilitate production predictions when out-of-range (unseen) input parameters have to extrapolate from data and that transferring the weights learned from the source field to the target field can add valuable information to enhance the prediction performance of the target field. Introduction While the limitations of simulation models make them unreliable for developing unconventional reservoirs, the increasing number of wells drilled in recent decades has provided an extensive database for tight oil horizontal wells. The use of such large data sets has permitted the use of data-driven proxy models. These models rely on estimating the approximate functional input/output relations by using various statistical and mathematical techniques. Neural networks are one such type of model designed to emulate neurons in the brain to find patterns within data that have complex relations. Implementation of data-driven proxy models has been used in the characterization of both unconventional and conventional reservoirs. However, these models cannot extrapolate data outside the training-data range. Transfer learning is one way in which a data-driven model trained on a large data set can be used to learn relevant and common features to solve a specific task with a given source data set. These features are then transferred to build a different model for a target data set with a similar problem definition. The method proposed by the authors attempts to take the benefits of both physics-based simulators and data-driven proxy models to build a more-robust model that can be transferred to build a reliable proxy model for a new field with limited data. In the model presented in the complete paper, all discrepancies between the simulated responses and field production data can be compensated. Then, the model is transferred to a second neural network designed to be a proxy model for the target field with much smaller data. The neural-network model will attempt to capture the correlations between its formation properties and the difference in production responses from this field. The equations vital to this process are provided in the complete paper.