GHGEAT: Gibbs-Helmholtz-Constrained Graph External Attention for Temperature-Dependent Activity Coefficient...

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Thermodynamics

Background:

• Activity coefficients are crucial for chemical process design, measuring real solution deviations from ideal behavior.
• Existing models often lack thermodynamic consistency and struggle with accurate molecular representations.
• Addressing these gaps is vital for improving the reliability of chemical simulations.

Purpose of the Study:

• To introduce a novel model, Gibbs-Helmholtz Graph External Attention (GHGEAT), for predicting activity coefficients.
• To enhance thermodynamic consistency and molecular feature extraction in predictive models.
• To provide a physically meaningful decision-making tool for chemical engineering applications.

Main Methods:

• Developed GHGEAT, a model utilizing external attention for global molecular feature extraction.
• Integrated the Gibbs-Helmholtz equation to enforce rigorous physical constraints.
• Validated GHGEAT on internal and a new large-scale external dataset (IDAC_2026).

Main Results:

• GHGEAT achieved a Mean Absolute Error (MAE) of 0.07 on an internal dataset (21,048 points).
• On the out-of-distribution IDAC_2026 dataset, GHGEAT achieved an MAE of 0.3407 and an R² of 0.9343.
• Demonstrated superior performance over benchmark models, with over 0.25 absolute R² improvement.

Conclusions:

• GHGEAT combines hierarchical feature extraction with thermodynamic consistency for robust predictions.
• The model shows strong performance in temperature interpolation and extrapolation.
• GHGEAT offers a physically meaningful tool for critical chemical engineering applications.