Development and Application of an Artificial Neural Network Tool for Chemical EOR Field Implementations

Abstract

Abstract The field-scale design of chemical enhanced oil recovery (cEOR) processes requires running complex numerical models that are computationally demanding. This paper provides an efficient screening platform for the cEOR feasibility study by presenting five artificial neural network (ANN) based models. We constructed 1,100 ANN training cases using CMG-STARS to capture the variation in reservoir petrophysical properties and the range of injected chemicals properties for a five-spot pattern. The design parameters were coupled with the reservoir properties using several functional links to optimize the ANN models and improve their performances. The training cases were employed using back-propagation methods to construct one forward model (Model #1) and four inverse models. Model #1 predicts reservoir response (i.e., oil rate, water cut, injector bottomhole pressure, cumulative oil) for known reservoir characteristics (i.e., permeability, thickness, residual oil saturation, chemical adsorption) and project design parameters (i.e., pattern size, chemical slug size and concentration), Model #2 predicts reservoir characteristics by history matching the reservoir response, and Model #3 predicts project design parameters for known reservoir response and characteristics. Models #4 and #5 predict project design parameters for a targeted cumulative oil volume and project duration time, which is useful for economical evaluation before the implementation of cEOR projects. The validation results show that the developed ANN-based models closely predict the numerical results. In addition, the models are able to reduce the computational time by four orders of magnitude, which is significant considering the complexity of cEOR modeling and the need for reliable and efficient tools in building cEOR feasibility studies. In terms of accuracy, Model #1 has a prediction error of 5% whereas the error for other four inverse ANN models is about 20–40%. To enhance the performance of the inverse ANN models, we changed the ANN structure, increased training cases, and used functional links, which slightly reduced the error. Further, we introduced a back-check loop that uses the predicted parameters from the inverse ANN models as inputs in the forward ANN model. A comparison of back-check results for the reservoir response with the numerical results delivers a relatively small error of 10%, revealing the non-uniqueness of solutions obtained from the inverse ANN models.