2025 · CVPR 2025 — MM4Mat: Multimodal Learning for Materials Science workshop · Preprint
Understanding the Capabilities of Molecular Graph Neural Networks in Materials Science Through Multimodal Learning and Physical Context Encoding
Can Polat , Hasan Kurban , Erchin Serpedin , Mustafa KurbanResearch highlights
What chemical context adds to molecular GNNs
Test whether textual chemical descriptors improve predictions made from molecular geometry.
Why we investigated it
XYZ coordinates omit information available in chemical databases, but extra descriptors may help some targets more than others.
What the work contributes
A gated fusion framework combines molecular graphs with textual descriptors such as names, formulas, and physicochemical properties.
What the evidence shows
Adding text improves certain electronic-property predictions while gains remain limited for other targets. Compared GNNs show similar patterns of success and failure.
Build on this work
Use the paper's modality comparisons to test whether added descriptors contribute useful information for your target properties.
- Molecular graph neural networks
- Multimodal learning
- Chemical descriptors
- Physical context
Abstract
Molecular graph neural networks (GNNs) often focus exclusively on XYZ-based geometric representations and thus overlook valuable chemical context available in public databases like PubChem. This work introduces a multimodal framework that integrates textual descriptors, such as IUPAC names, molecular formulas, physicochemical properties, and synonyms, alongside molecular graphs. A gated fusion mechanism balances geometric and textual features, allowing models to exploit complementary information. Experiments on benchmark datasets indicate that adding textual data yields notable improvements for certain electronic properties, while gains remain limited for others. Furthermore, the GNN architectures display similar performance patterns (improving and deteriorating on analogous targets), suggesting they learn comparable representations rather than distinctly different physical insights.
Cite
@misc{polat2025understanding,
title = {Understanding the Capabilities of Molecular Graph Neural Networks in Materials Science Through Multimodal Learning and Physical Context Encoding},
author = {Can Polat and Hasan Kurban and Erchin Serpedin and Mustafa Kurban},
year = 2025,
journal = {CVPR 2025 — MM4Mat: Multimodal Learning for Materials Science workshop},
doi = {10.48550/arxiv.2505.12137},
url = {https://arxiv.org/pdf/2505.12137},
}