Gibbs–Helmholtz graph neural network: capturing the temperature dependency of activity coefficients at infinite dilution

The accurate prediction of physicochemical properties chemical compounds in mixtures (such as the activity coefficient at infinite dilution $\gamma_{ij}^\infty$) is essential for developing novel and more sustainable processes. In this work, we analyze performance previously-proposed GNN-based models $\gamma_{ij}^\infty$, compare them with several mechanistic a series 9 isothermal studies. Moreover, develop Gibbs-Helmholtz Graph Neural Network (GH-GNN) model predicting $\ln \gamma_{ij}^\infty$ molecular systems different temperatures. Our method combines simplicity Gibbs-Helmholtz-derived expression graph neural networks that incorporate explicit intermolecular descriptors capturing dispersion hydrogen bonding effects. We have trained using experimentally determined data 40,219 binary-systems involving 1032 solutes 866 solvents, overall showing superior compared to popular UNIFAC-Dortmund model. GH-GNN continuous discrete inter/extrapolation give indications model's applicability domain expected accuracy. general, able produce predictions extrapolated if least 25 same combination solute-solvent classes are contained training set similarity indicator above 0.35 also present. This its recommendations been made open-source https://github.com/edgarsmdn/GH-GNN.