Artificial neural network modeling of scale-dependent dynamic capillary pressure effects in two-phase flow in porous media

Abstract

A computationally efficient and simple alternative platform for the prediction of the domain scale dependence of the dynamic capillary pressure effects, defined in terms of a coefficient named as dynamic coefficient (τ), is developed using an artificial neural network (ANN). The input parameters consist of the phase saturation, media permeability, capillary entry pressure, viscosity ratio, density ratio, temperature, pore size distribution index, porosity and domain volume with corresponding output τ obtained at different domain scales. Different ANN configurations as well as linear and nonlinear multivariate regression models were tested using a number of performance criteria. Findings in this work showed that the ANN structures with double hidden layers perform better than those with a single hidden layer. In particular, the ANN configuration with 13 and 15 neurons in the first and second hidden layers, respectively, performed the best. Using this best-performing ANN, effects of increased domain size were predicted for three separate experimental results obtained from literature and our laboratory with different domain scales. Results showed increased magnitude of τ as the domain size increases for all the independent experimental data considered. This work shows the applicability and techniques of using ANNs in the prediction of scale dependence of two-phase flow parameters.