Quantum-Chemically Informed Representation Learning for Molecular Crystal Property Prediction via Multimodal Graph Neural Networks
Machine-learning-based modeling of molecular crystals is limited by representations that encode geometric structure while overlooking the electronic features that often govern solid-state behavior. We introduce a multimodal learning framework that integrates crystal graphs with solid-state Quantum Theory of Atoms in Molecules (QTAIM)-informed molecular graphs for enhanced predictions of solid-state properties. Our architecture employs modality-specific graph neural network encoders and a bidirectional gated fusion mechanism to capture complementary information between the two representations. When applied to band gap prediction in organic molecular crystals, the fused model outperforms geometry-only and molecular-only baselines. Analysis of the learned representations reveals that QTAIM descriptors contribute interpretable signals associated with intermolecular interactions and density redistribution within the crystal, highlighting the value of quantum-informed learning for advancing data-driven solid-state modeling.