Physics-Embedded Machine Learning Model for Phase Equilibrium Prediction in Multicomponent Systems.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Predicting activity coefficients in liquid mixtures is crucial for chemical process design.
• Existing models like COSMO-SAC rely on complex quantum chemistry calculations.
• A need exists for accurate and scalable methods using simpler molecular representations.

Purpose of the Study:

• To develop a novel machine learning framework, TeNNet-SAC, for predicting activity coefficients.
• To utilize only SMILES representations for input, simplifying data requirements.
• To achieve high accuracy and thermodynamic consistency in predictions.

Main Methods:

• TeNNet-SAC integrates a σ-profile predictor, a geometry predictor, and a Γ predictor.
• Predictors are trained on quantum solvation calculations and synthetic data.
• The model is fine-tuned with experimental activity coefficient data for enhanced accuracy.

Main Results:

• The base TeNNet-SAC model shows accuracy comparable to COSMO-SAC.
• The fine-tuned TeNNet-SAC model consistently outperforms COSMO-SAC.
• The model demonstrates natural generalization to multicomponent mixtures.

Conclusions:

• TeNNet-SAC offers a robust and scalable alternative for activity coefficient prediction.
• The framework leverages machine learning for efficient chemical property estimation.
• The approach satisfies thermodynamic consistency, ensuring reliable predictions.