Application of Robust Neural Networks to Modeling Crude Oil Viscosity Reduction Using Nanoparticles

Viscosity plays a critical role in determining the flow behavior and recoverability of crude oil, particularly in enhanced oil recovery (EOR) processes. Recent advancements in nanotechnology have introduced nanoparticles (NPs) as effective agents for reducing oil viscosity and improving EOR efficiency. In this study, 646 experimental data points were used to model the reduction in crude oil viscosity achieved through the NPs treatment. Four neural network architectures were applied to predict the final-to-initial viscosity ratio: multilayer perceptron (MLP), cascade-forward neural network (CFNN), radial basis function network (RBF), and generalized regression neural network (GRNN). Among them, the CFNN trained using the Levenberg-Marquardt algorithm (CFNN-LM) demonstrated the highest performance, achieving the lowest overall average absolute relative error of 1.83% and the highest coefficient of determination (R ²) of 0.9851, with the MLP-LM model closely following. In contrast, the GRNN and RBF models demonstrated weaker predictive capabilities characterized by higher errors and lower R ² values. The CFNN-LM model also effectively captured key physical trends, showing that viscosity reduction improves with increasing temperature, NPs concentration, and shear rate, while smaller NP sizes further enhance the effect. Moreover, Pearson sensitivity analysis identified initial oil viscosity (27%), NPs density (24%), and NPs size (22%) as the main linear factors, while Spearman sensitivity analysis emphasized NPs concentration (30%), initial viscosity (25%), and temperature (21%) as dominant nonlinear contributors. Overall, the intrinsic properties of crude oil and characteristics of NPs are the primary drivers of viscosity reduction, with concentration and temperature exerting stronger nonlinear effects. Finally, the leverage method confirmed the reliability of the model, identifying only three suspected and 47 out-of-leverage points, showing both the validity of the data set and the predictive robustness of the CFNN-LM model developed in this study.