Restricted (Penn State Only)
Abdullah, Mohammad B
Graduate Program:
Energy and Mineral Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
February 18, 2019
Committee Members:
  • Hamid Emami-Meybodi, Thesis Advisor
  • Turgay Ertekin, Thesis Advisor
  • Gregory R King, Committee Member
  • ANN
  • EOR
  • cEOR
  • Enhanced Oil Recovery
  • Machine Learning
  • Artificial Neural Network
  • Field Implementation
Field-scale design of chemical enhanced oil recovery (cEOR) processes requires complex numerical models that are computationally expensive. This thesis presents an “artificial expert” tool, using an artificial neural network (ANN) based model as an efficient screening platform during a cEOR feasibility study. To optimize the ANN models and improve its performance, the design parameters were coupled with the reservoir properties using several functional links. A reservoir simulation model with average reservoir properties was constructed using CMG-STARS. Over 1000 ANN training cases were extracted from the reservoir model to capture variations in the reservoir petrophysical properties and properties of the range of injected chemicals. The training cases were employed to construct five ANN models using the back-propagation training method. To enhance ANN pattern recognition, functional links were used to signify the dominant factors in cEOR, including adsorption, residual resistance factor, and injected chemicals concentrations. After validating the ANN models, they were built into a user-friendly interface. In this study, five ANN models are proposed. The first model predicts the production profile for a given reservoir and project design, the second predicts reservoir characteristics by history matching the production profile, and the third predicts project design parameters for a given reservoir and production profile. The remaining two are specific to economic evaluation prior to project implementation by predicting the project design parameters for a targeted cumulative oil volume and project execution period. Predicting production profile showed a small prediction error of 5 percent using a “forward ANN model.” In contrast, the other four “inverse ANN models” produced prediction errors of between 20 and 40 percent. Different trials were tested to enhance ANN pattern recognition — changing the ANN structure, increasing training cases and using functional links — but the error rate did not drop below 20 percent. Hence, a back-check loop was created using the predicted “inverse ANN models” parameters as inputs into the “forward ANN model.” When comparing these results with the numerical production profile, the error was less than 10 percent. Clearly, the high prediction error in the “inverse ANN models” was due to the non-uniqueness of the solution rather than weak pattern recognition. In other words, there are different combinations of design parameters and reservoir characteristics that correspond to the same production profile. The constructed ANN production profile prediction error was within 5–10 percent compared to the results of the numerical simulation model, reducing computational time by up to four orders of magnitude. This is a significant breakthrough considering the complexity involved in cEOR modeling and the need for a reliable and efficient tool in building cEOR feasibility studies. This paper presents a novel method by testing over 30 functional links involving physical and mathematical representation of cEOR parameters to improve the effectiveness of the ANN’s pattern recognition capabilities.