Similarity-Informed Matrix Completion Method for Predicting Activity Coefficients.

Area of Science:

• Chemical Engineering
• Thermodynamics
• Machine Learning

Background:

• Accurate prediction of thermodynamic properties like activity coefficients is crucial for chemical process design.
• Physics-based methods have limitations in accuracy and scope.
• Machine learning, specifically matrix completion methods (MCMs), shows promise but struggles in data-sparse regions.

Purpose of the Study:

• To develop a hybrid matrix completion method (MCM) for predicting activity coefficients at infinite dilution at 298 K.
• To improve prediction accuracy in data-sparse regions by incorporating synthetic training data.
• To analyze the impact of different training data types on prediction performance.

Main Methods:

• A novel hybrid matrix completion method (MCM) was developed.
• The hybrid MCM integrates experimental data with synthetic data from modified UNIFAC (Dortmund) and a similarity-based approach.
• The performance was evaluated based on prediction accuracy, particularly in data-sparse regions.

Main Results:

• The hybrid MCM demonstrates robust performance, excelling in regions with limited data.
• Incorporating synthetic data from modified UNIFAC (Dortmund) and the similarity-based approach significantly improves MCM performance when experimental data are sparse.
• High accuracy requires training sets that include mixtures similar to those being predicted, even with abundant experimental data.

Conclusions:

• The proposed hybrid MCM offers a more robust prediction framework for activity coefficients compared to traditional methods.
• Synthetic data plays a critical role in enhancing MCM performance, especially in data-limited scenarios.
• The composition of the training data, including similarity to target mixtures, is vital for achieving high prediction accuracy.