Okon John*, Tinuola Udoh, Blessed Emenka
Reducing the Interfacial Tension (IFT) between crude oil and brine is one of surfactants' key functions in Enhanced Oil Recovery (EOR). Surfactants improve the mobility and displacement of oil by water or other fluids by lowering the IFT. This makes it easier to release and mobilize oil that has been trapped and would otherwise be challenging to recover. The conventional methods used to assess whether surfactants are effective in lowering the IFT of crude oil and brine in a reservoir entail a number of costly, time-consuming, and difficult processes. These include a thorough examination of the reservoir's characteristics, an analysis of the composition of the crude oil, a study of the surfactant's properties, and a battery of extensive laboratory tests to ascertain the surfactant's efficacy in lowering the IFT and enhancing oil recovery. These difficulties will be resolved by using Machine Learning (ML) techniques to artificially intelligently predict crude oil-brine IFT based on surfactant properties. Machine learning, a branch of artificial intelligence, is essentially the use of computer algorithms to predict the future with (supervised learning) or without (unsupervised learning) prior knowledge of the past. In order to forecast crude oil-brine IFT using surfactant properties as dependent variables, this work concentrated on developing a high-level ensemble machine learning model based on the "boosting" algorithms, namely the Gradient Boosting Decision Tree (GBDT) and the Adaptive Boosting (ADABOOST) algorithms. Four models were created, two for each algorithm, depending on the base learner and the quantity of dependent variables. The models were trained, tested, and assessed to identify the optimal model after being fitted with surfactants and crude oil-brine IFT data. The impact and effects of training the models with different data sizes, functional forms, and decision-making processes to predict are investigated in the early stages of the simulation. As is recommended for predictive machine learning models, the models were then assessed using the statistical metrics of Root Mean Squared Error (RMSE), coefficient of determination (R2), Standard Deviation (SD), and Average Absolute Relative Deviation (AARD). The GBDT model-2 performed the best out of the four developed models, according to the evaluation results, with an R2 value of 99.70%, an RMSE of 0.103, an AARD of 1.32%, and an SD value of 0.0327.
Published Date: 2025-01-10; Received Date: 2024-03-26
