HANNA: hard-constraint neural network for consistent activity coefficient prediction.

Area of Science:

• Thermodynamics and Physical Chemistry
• Computational Chemistry and Materials Science
• Chemical Engineering

Background:

• Activity coefficients are crucial thermodynamic properties for mixture behavior in science and engineering.
• Traditional neural networks often fail to enforce physical laws, leading to inconsistent predictions.
• Existing models like UNIFAC have limitations in accuracy and applicability.

Purpose of the Study:

• To introduce the first hard-constraint neural network model (HANNA) for predicting activity coefficients.
• To ensure thermodynamic consistency by embedding physical laws directly into the model architecture.
• To develop a universally applicable model requiring only component SMILES as input.

Main Methods:

• Developed a deep-set neural network architecture incorporating hard constraints for thermodynamic consistency.
• Ensured symmetry and adherence to the Gibbs-Duhem equation within the model.
• Trained and validated the model using over 317,000 data points from the Dortmund Data Bank.

Main Results:

• HANNA achieved significantly higher prediction accuracy for activity coefficients compared to the state-of-the-art UNIFAC model.
• The model demonstrated robustness and consistency by adhering to fundamental thermodynamic principles.
• Inputting only SMILES strings enabled application to any binary mixture.

Conclusions:

• The HANNA model represents a breakthrough in predicting thermodynamic mixture properties with physical rigor.
• Hard-constraint neural networks offer a superior approach to modeling chemical systems.
• HANNA is an open-source, highly accurate, and broadly applicable tool for researchers and engineers.