Predicting Newtonian viscosity using machine learning trained on equilibrium molecular dynamics.

Area of Science:

• Computational chemistry
• Rheology
• Materials science

Background:

• Newtonian viscosity (η0) is fundamental to liquid rheology and constitutive modeling.
• Equilibrium molecular dynamics (EMD) via Green-Kubo formalism calculates η0 but demands extensive simulation times.
• Accurate stress autocorrelation evaluation in EMD is computationally prohibitive.

Purpose of the Study:

• To integrate EMD with machine learning for modeling the relationship between η0 and stress correlation time (τ).
• To develop a computationally efficient framework for determining equilibrium viscosity.
• To support the development of rate-dependent viscosity models.

Main Methods:

• Utilized equilibrium molecular dynamics (EMD) simulations for liquid n-hexadecane.
• Developed and trained a three-layer artificial neural network on EMD data.
• Modeled the relationship between Newtonian viscosity (η0) and stress correlation time (τ).

Main Results:

• The machine learning model accurately captures the onset and evolution of the viscosity plateau under varying thermodynamic conditions.
• The model successfully reproduces direct EMD results.
• The model aligns with extrapolated non-equilibrium predictions, demonstrating physical consistency.

Conclusions:

• The integration of EMD and machine learning provides a physically consistent and computationally efficient method for determining equilibrium viscosity.
• This approach overcomes the limitations of long simulation times in traditional EMD.
• The developed framework supports advancements in constitutive modeling and rate-dependent viscosity predictions.