Temperature-Dependent Density and Viscosity Prediction for Hydrocarbons: Machine Learning and Molecular Dynamics...

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Lubricant formulation relies on base oils, which are complex hydrocarbons.
• Predicting material properties like density and viscosity is crucial for lubricant design.
• Traditional methods for property prediction can be time-consuming and resource-intensive.

Purpose of the Study:

• To develop accurate machine learning models for predicting temperature-dependent density and dynamic viscosity of complex hydrocarbons.
• To explore the utility of dynamic descriptors from molecular dynamics simulations for property prediction.
• To identify key molecular descriptors influencing hydrocarbon properties.

Main Methods:

• Gaussian process regression (GPR) models were employed.
• Predictor selection utilized LASSO regularization and domain knowledge.
• Hyperparameter optimization was performed using Bayesian optimization.
• Molecular dynamics simulations generated dynamic descriptors.

Main Results:

• GPR models achieved high accuracy, with R² values of 99.6% for density and 97.7% for viscosity.
• Models using dynamic descriptors performed comparably to those using numerous static descriptors.
• Model interpretability techniques (PDP, ICE, LIME) identified important static and dynamic predictors.

Conclusions:

• Quantitative structure-property relationship (QSPR) models provide versatile predictions for hydrocarbon density and viscosity.
• Dynamic simulation descriptors are effective for predicting material properties.
• The study offers a robust framework for materials design and property prediction.