DeepThermoMix: a local composition graph neural networks model for multicomponent activity coefficients

We propose DeepThermoMix, a local-composition-based graph neural network model that predicts multicomponent activity coefficients directly from molecular graphs. Unlike classical local composition models, which rely on fixed algebraic expressions, DeepThermoMix generalizes the local composition assumption by learning interaction patterns without assuming predefined functional forms, such as the Boltzmann distribution. Unlike existing constrained machine learning models that remain limited by the pairwise additivity assumption, DeepThermoMix can capture emergent N-body interactions while preserving linear scaling with the number of components, thereby avoiding the combinatorial explosion inherent to pairwise approaches. Furthermore, thermodynamic consistency is directly embedded into the architecture by predicting the molar excess Gibbs free energy as a scalar potential from which activity coefficients are derived via automatic differentiation. Despite being trained on quantum-based COSMO-RS data, DeepThermoMix naturally converges toward surfaces resembling classical formalisms, a phenomenon indicative of "structural regularization." This study positions DeepThermoMix as a thermodynamically consistent and scalable model for multicomponent activity coefficient prediction.