Making thermodynamic models of mixtures predictive by machine learning: matrix completion of pair interactions

Area of Science:

• Thermodynamics
• Chemical Engineering
• Machine Learning

Background:

• Predictive models for thermodynamic properties of mixtures are crucial in chemical engineering and chemistry.
• Classical thermodynamic models excel at continuous variations (temperature, concentration).
• Matrix completion methods (MCMs) from machine learning generalize well across discrete binary systems.

Purpose of the Study:

• To combine the strengths of classical thermodynamic models and MCMs for improved mixture property prediction.
• To develop a hybrid model that predicts pair-interaction energies using MCMs.
• To create a complete set of parameters for the UNIQUAC model applicable to a vast number of components.

Main Methods:

• Embedding a matrix completion method (MCM) into the UNIQUAC thermodynamic model.
• Training the hybrid model using a Bayesian machine-learning framework.
• Utilizing experimental activity coefficient data for binary systems from the Dortmund Data Bank (1146 components).

Main Results:

• Achieved a complete set of UNIQUAC parameters for all binary systems of the studied components.
• Enabled prediction of activity coefficients at arbitrary temperatures and compositions for binary and multicomponent systems.
• Demonstrated superior performance compared to the modified UNIFAC (Dortmund) model in predicting activity coefficients.

Conclusions:

• The hybrid MCM-UNIQUAC model offers a powerful new approach for predicting thermodynamic properties of liquid mixtures.
• This method provides accurate predictions across a wide range of conditions and component combinations.
• The developed model surpasses current state-of-the-art physical models, advancing the field of mixture property prediction.