A data-driven study on viscosity estimation of hydrogen-containing gas mixtures using machine learning

Abstract

Hydrogen has increasingly gained attention as an energy carrier due to its clean nature and high efficiency. One crucial factor in its production, storage, and transportation is viscosity. However, none of the previous studies have explored the use of machine learning models to estimate the viscosity of hydrogen-based mixtures using laboratory data. To address this gap, we compiled a dataset of 1624 viscosity measurements for hydrogen-based gas mixtures from past experimental studies. Six machine learning techniques were employed for modeling: k-nearest neighbors (KNN), support vector regression (SVR), regression tree, categorical boosting (CatBoost), extra trees, and extreme gradient boosting (XGBoost). The dataset was split into 70 % for training and 30 % for testing, with a 5-fold cross-validation approach applied to validate the models during training. To assess model performance, we used cross plots, residual error plots, error metrics, and absolute error frequency plots. Among all methods, the extra trees model demonstrated the highest accuracy, achieving an R² value of 0.9983. It was followed closely by XGBoost (0.9976), CatBoost (0.9974), KNN (0.9923), regression tree (0.9917), and SVR (0.9735). Sensitivity analysis revealed that temperature had the most significant impact on viscosity, whereas methane mole fraction had the least. Additionally, at low pressures, the mole fractions of carbon dioxide and methane exhibited an inverse relationship with viscosity, while the hydrogen mole fraction showed a direct correlation. To define the applicability range of the extra trees model, a William's plot was used, indicating that 1562 data points (96 % of the dataset) were valid. Given the direct impact of viscosity on flow behavior and system efficiency, these findings can be instrumental in optimizing hydrogen production, transportation, and storage processes.