Multi-modal foundation model for materials science. Modality-specific foundation models encode the representations to the latent space where latent vectors are fused, so that a shared Platonic representation of molecules/materials that unifies all material representations can be learned.
Image: Changwen Xu, University of Michigan and Hongshuo Huang, University of Michigan
This project will develop multimodal AI foundation models to rapidly predict and design new materials, accelerating discovery across applications like energy storage and electronics by efficiently screening millions of candidate materials.
The development of new materials is fundamental to technological progress, from electronics and medicine to clean energy and aerospace. However, traditional materials discovery is painfully slow—often taking decades to move from laboratory to real- world application. This project will develop breakthrough artificial intelligence models that can rapidly identify and design new materials across diverse applications. The team's "multimodal" foundation models will simultaneously process multiple types of information about materials creating a comprehensive understanding of how materials behave. The multimodal foundation model for materials will act as a powerful accelerator for materials discovery, enabling researchers to screen millions of potential materials. The models will be trained using Aurora and will support diverse materials discovery challenges in domains such as energy storage and electronics.
