Non-linear AI based viscosity prediction for binary brine data

Abstract

A semi-empirical framework is developed to estimate the viscosity of binary brine systems (m-Pas) over temperatures from 293 to 323 K and molal concentrations between 0 and 4, using a continuous viscosity function η[m,T] constructed from experimental salt-water data. The modeling employs Machine Learning (ML) algorithms – Artificial Neural Network (ANN), Random Forest Regressor (RFR), and Gaussian Process Regressor (GPR) to optimize prediction accuracy with fewer fitting parameters. Comparative assessment shows that the AI-based approach achieves an average absolute deviation (%AAD) below 1%, outperforming conventional models while utilizing fewer parameters. Specifically, the GPR model yielded the best results with a %AAD of 0.96%, root mean square error (RMSE) of 0.0001, and an R2 score of 0.9979, surpassing RFR (%AAD 6.65%) and ANN (%AAD 18.67%). Robustness and model validity are confirmed through rigorous cross-validation and confidence intervals. These findings demonstrate that machine learning, particularly GPR, provides a highly accurate and practical approach for predicting physicochemical properties in aqueous electrolyte environments, enabling effective industrial applications in process optimization and resource management. The results underscore machine learning as a practical approach for modelling physicochemical properties within aqueous electrolyte environments.