Multimodal fusion of complementary material representations for generalizable and interpretable property prediction

Material properties often depend on coupled compositional, structural, and geometric factors that are incompletely captured by any single representation. We introduce interpretable multimodal fusion (IM-Fuse), a framework that integrates compositional descriptors, crystal graphs, and radial distribution functions for material property prediction. Using 10,123 ion-insertion reactions as a representative materials dataset, we find that fusion does not uniformly improve random-split accuracy but improves generalization stability under chemistry-aware out-of-distribution evaluations. Attribution and interaction analyses show how composition, local coordination, and radial packing jointly shape predictions through structured relationships across representations. These results establish multimodal fusion as a general framework for auditing representation complementarity in data-driven materials informatics.