Data-Driven Modeling of Low-Salinity Effects on Multiphase Flow Using a Dual Artificial Neural Network Scheme

Summary

An experimental database of relative permeability tests for low-salinity water injection (LSWI) and engineered water injection (EWI) is analyzed in terms of rock/fluid properties and used for the development of data-driven artificial neural network (ANN) models for relative permeability prediction under salinity variation. The exploratory analysis of the compiled database confirmed salinity-related trends reported in the literature, such as the reduction in residual oil saturation with decreasing salinity. A computed wettability index also indicated a shift toward more water-wet or intermediate conditions with salinity reduction, more notably in carbonate samples. Using these data, a unified machine learning framework is proposed to directly predict low-salinity (LS) relative permeability curves from high-salinity (HS) relative permeability curves and salinity conditions, without relying on additional coreflood experiments or predefined empirical correlations. For this purpose, a multilayer feedforward architecture with backpropagation is used in a dual-network scheme: One network predicts the normalized coordinate points of the relative permeability curves, while the other estimates the endpoints, using an appropriate subset of the data. This dual-network strategy separates the learning of curve shape and endpoint variations induced by salinity reduction, mitigating issues of endpoint overestimation observed in previous models. The trained models achieved a good performance in cross-validation and produced more stable predictions. The proposed approach offers a practical tool for estimating relative permeability in LSWI/EWI projects. Furthermore, the methodology is directly applicable to related problems, such as immiscible displacements in carbon capture, utilization, and storage (CCUS), and to other multivariate regression tasks.