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

Activity coefficients quantify the deviation of a real solution from an ideal one and are essential for chemical process design. To address the structural representation gaps and lack of thermodynamic consistency in existing models, we propose Gibbs-Helmholtz Graph External Attention (GHGEAT), a novel model that extracts global molecular features via an external attention mechanism and integrates the Gibbs-Helmholtz equation to establish rigorous physical constraints. GHGEAT achieves a Mean Absolute Error (MAE) of 0.07 on a public internal data set of 21,048 data points, outperforming established models. To further evaluate generalization capability, GHGEAT was tested on the IDAC_2026 data set, a new large-scale external data set constructed in this study by rigorously filtering the Brouwer data set to exclude training overlaps. On this challenging out-of-distribution data set, GHGEAT demonstrates superior performance with an MAE of 0.3407 and an R² of 0.9343, representing an absolute R² improvement of over 0.25 compared to second-best benchmark models. By combining hierarchical global feature extraction with thermodynamic consistency, GHGEAT ensures robust performance in continuous temperature interpolation and extrapolation, providing a physically meaningful decision-making tool for critical chemical engineering applications. The source code is available at https://github.com/Wang213-wq/GHGEAT.