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

Local composition theory postulates that the immediate molecular environment of a component differs from the bulk mixture composition. Classical local composition models describe this by using fixed algebraic expressions in the form of Boltzmann-like distributions which in turn results in explicit pairwise enumeration of components in multicomponent mixtures. Recent hard-constrained neural networks relax this functional form but remain tied to pairwise enumeration through geometric projection. Here, we propose DeepThermoMix, a local-composition-based graph neural network that predicts multicomponent activity coefficients directly from molecular graphs. Our approach bypasses this by processing the mixture as a whole, allowing training on datasets with varying mixture sizes, and yielding linear computational scaling as the number of components grows. The excess Gibbs energy is itself constructed as a departure from the ideal solution, with the reference state represented by the network rather than fixed by a prescribed algebraic factor. Furthermore, thermodynamic consistency is guaranteed as a consequence of predicting the molar excess Gibbs energy as a scalar potential from which the activity coefficients are derived via automatic differentiation. Notably, despite being trained on quantum-based COSMO-RS data, the model seems to exhibit selective tendencies to diverge from its training data characteristics toward a particular thermodynamic formalism, a phenomenon which we term "structural regularization". Building upon the recent shift toward introducing thermodynamic rigor within flexible neural networks, this work retains the physical statement of local composition theory while relaxing its Boltzmann-like distribution form. By departing from the pairwise additivity assumption, we establish a scalable and thermodynamically rigorous approach for multicomponent systems, with the present work laying an isothermal foundation for future temperature-dependent extensions.