Thermodynamics-consistent graph neural networks

Area of Science:

• Thermodynamics
• Machine Learning
• Chemical Engineering

Background:

• Predicting activity coefficients is essential for chemical process design.
• Existing models often face limitations in thermodynamic consistency and applicability.
• Accurate prediction of composition-dependent properties is a key challenge.

Purpose of the Study:

• To introduce a novel graph neural network (GE-NN) approach for predicting activity coefficients.
• To ensure thermodynamic consistency in predictions using fundamental thermodynamic principles.
• To develop a model free from thermodynamic modeling limitations.

Main Methods:

• Developed excess Gibbs free energy graph neural networks (GE-GNNs).
• Utilized automatic differentiation for end-to-end learning of activity coefficients.
• Ensured thermodynamic consistency by predicting molar excess Gibbs free energy.

Main Results:

• Achieved high accuracy in predicting activity coefficients for binary mixtures.
• Demonstrated inherent thermodynamic consistency without additional loss terms.
• The GE-GNN model showed no thermodynamic modeling limitations.

Conclusions:

• GE-GNNs offer a powerful and thermodynamically consistent method for predicting mixture properties.
• This approach advances the application of machine learning in chemical thermodynamics.
• The model provides reliable predictions essential for chemical process optimization.