Predicting the properties of mixtures is crucial across chemical engineering, energy materials, environmental science, and biomedicine. It elucidates complex intercomponent interactions, improves experimental efficiency, and guides the design and optimization of new systems. However, traditional computational chemistry methods cannot keep pace with modern high-throughput workflows, and existing mixture prediction algorithms do not sufficiently capture intercomponent interactions and lack sufficient transferability, hindering practical adoption and broader dissemination. To address these challenges, we propose a multi-level cross-molecular interaction neural network (MCMINet), a novel native mixture modeling architecture with end-to-end component permutation invariance and generality. The model explicitly captures intercomponent interactions via cross-molecular interaction mechanisms and fuses structural information at both the molecular and mixture levels to construct mixture representations. This design preserves intrinsic information while enhancing interaction information, thereby improving predictive accuracy. Furthermore, inspired by the consistency between statistical mechanics and thermodynamics, we design a loss function ℒ mix for mixture modeling that embeds composition dependence into a regularization term to enforce physical consistency during training and improve generalization. MCMINet demonstrates state-of-the-art performance in predicting electrolyte ionic conductivity, activity coefficients of non-ideal solutions, and vapor–liquid equilibrium behavior. Experiments spanning diverse scales, mixing ratios, component counts, and physicochemical properties enable researchers to gain deeper insights into the non-additivity of intercomponent interactions, offering a new perspective on “structure–property” research in mixtures. This approach holds significant potential for downstream tasks such as process optimization, chemical screening, and drug formulation.
Shi et al. (Sat,) studied this question.