Machine Learning in Thermodynamics: Prediction of Activity Coefficients by Matrix Completion.

Area of Science:

• Chemical Engineering
• Physical Chemistry
• Computational Chemistry

Background:

• Activity coefficients quantify liquid mixture nonideality, crucial for chemical engineering.
• Accurate prediction is needed for unexplored binary mixtures.
• Current prediction methods require extensive data and refinement.

Purpose of the Study:

• To propose a novel probabilistic matrix factorization model for predicting activity coefficients.
• To demonstrate superior performance compared to existing state-of-the-art methods.
• To enable accurate prediction for a wider range of binary mixtures.

Main Methods:

• Probabilistic matrix factorization model.
• Utilized existing experimental activity coefficient data.
• No physical component descriptors were incorporated.

Main Results:

• The proposed model accurately predicts activity coefficients in arbitrary binary mixtures.
• Outperformed the state-of-the-art method in predictive accuracy.
• Required significantly less training effort than existing methods.

Conclusions:

• The probabilistic matrix factorization model offers a powerful new approach for predicting physicochemical properties.
• This method has the potential to revolutionize modeling and simulation in chemical engineering.
• Enables accurate predictions for previously inaccessible binary mixtures.