Thermodynamically consistent machine learning model for excess Gibbs energy

Area of Science:

• Thermodynamics
• Chemical Engineering
• Machine Learning

Background:

• Excess Gibbs energy is crucial for modeling thermodynamic properties of liquid mixtures.
• Predicting this property for multi-component mixtures from molecular structures remains a significant challenge.

Purpose of the Study:

• To develop a machine learning model, HANNA, for predicting excess Gibbs energy.
• To ensure thermodynamically consistent predictions by integrating physical laws as hard constraints.

Main Methods:

• Developed HANNA, a flexible machine learning model.
• Trained HANNA on experimental data for vapor-liquid equilibria, liquid-liquid equilibria, activity coefficients at infinite dilution, and excess enthalpies.
• Utilized a surrogate solver for end-to-end training on liquid-liquid equilibrium data.
• Employed a geometric projection method for extrapolation to multi-component mixtures.

Main Results:

• HANNA provides accurate predictions of excess Gibbs energy.
• The model demonstrates a substantially broader domain of applicability compared to state-of-the-art methods.
• Thermodynamically consistent predictions are guaranteed.

Conclusions:

• HANNA successfully addresses the challenge of predicting excess Gibbs energy for multi-component mixtures.
• The model offers improved accuracy and applicability in thermodynamic property prediction.
• Open-source code and an interactive interface are available for broader use.