A Deep Learning Approach for Screening Enhanced Oil Recovery Methods

With the increasing demand to meet the world’s energy needs and the high costs associated with Enhanced Oil Recovery projects, E&P companies are turning to machine learning to cut operating expenditures and development costs while optimizing hydrocarbon production. This research project aims to evaluate the ability of using neural networks, a subfield of machine learning, to screen for and select the most promising EOR technique among the available ones based on past successful EOR projects. A multilayer-perception neural network (MLP) with backpropagation was built and then trained with over 100 past EOR projects in order to screen for the best EOR technique. To optimize the model accuracy, we tested various parameters of the neural network to find the best model architecture: number of hidden layers, total number of neurons, activation function, and number of training epochs. After testing each of those sections, we came up with a chart with ranges that optimizes each parameter and yields the best model architecture for the neural network. Although this research project lays the foundations of incorporating deep learning into EOR projects and creates a robust model, more data gathering with an increasing number of input features (wettability, reservoir heterogeneity, injected fluid parameters) can lead to a more complex model with greater accuracy.