Multimodal artificial intelligence for inverse materials design

Traditional materials discovery requires screening vast numbers of compositions and structures to find materials with desired properties. This is computationally costly, while most machine-learning models rely on limited data sources, reducing their ability to efficiently explore the materials landscape.

To overcome these limitations, Gian-Marco Rignanese (WEL Research Institute – UCLouvain) and collaborators developed MEIDNet (Multimodal Equivariant Inverse Design Network), an inverse-design AI framework that starts from target properties and directly generates promising new materials. By combining structural and physico-chemical information in a shared representation space, MEIDNet improved learning efficiency and produced candidates that closely matched desired properties. These materials were subsequently validated through ab initio calculations, demonstrating their physical plausibility. The approach offers a powerful route to faster AI-assisted materials discovery for applications across a wide range of fields, including energy, electronics, and catalysis.

Reference: Babu et al, MEIDNet: multimodal generative AI framework for inverse materials design, npj Comput Mater (2026) https://doi.org/10.1038/s41524-026-02153-3