MixUni: Graph Foundation Modeling for Multi-Property Prediction of Molecular Mixtures

Accurate prediction of physico-chemical properties in multi-component liquid mixtures is essential for the discovery of next-generation materials. Existing approaches typically train separate models for each property and struggle to transfer across heterogeneous mixture datasets and physicochemical regimes. Moreover, conventional random-split evaluation often overestimates chemical generalisation by placing identical mixtures in both train and test sets. To address these limitations, we propose MixUni, a unified graph foundation modeling framework for joint multi-property prediction in chemical mixtures. MixUni combines an E(3)-equivariant molecular encoder with property-conditioned prediction heads, enabling a shared transferable representation across transport, thermodynamic, and structural property tasks while preserving property-specific inductive biases. In particular, MixUni incorporates a stretched-Arrhenius conductivity head for temperature-dependent ion transport and a sparse Mixture-of-Experts readout to adapt across heterogeneous property regimes. Experiments on five CheMixHub benchmarks show that MixUni consistently outperforms single-property geometric baselines, with especially strong gains under chemically out-of-distribution evaluation. These results demonstrate that unified graph representation learning combined with physics-guided readouts provides an effective foundation-modeling framework for reliable prediction in complex liquid mixtures.