Reproduce “Multimodal Foundation Models for Material Property Prediction and Discovery” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Viggo Moro, Charlotte Loh, Rumen Dangovski, Ali Ghorashi, Andrew Ma, Zhuo Chen, Samuel Kim, Peter Y. Lu and 2 more

Researchers from MIT and collaborators introduced MultiMat, a framework for self-supervised multimodal training of foundation models that integrates diverse material properties into a shared latent space. The system improves crystal property prediction by up to 10% Mean Absolute Error over state-of-the-art methods and facilitates material discovery through latent space similarity, revealing interpretable emergent physical characteristics.

Key insights

  • Integrating diverse material representations through multimodal pre-training enables the learning of more effective and transferable material representations for downstream tasks.
  • Aligning the latent spaces of different material modalities allows for a novel, rapid material discovery strategy based on cross-modality similarity retrieval.
  • The shared latent space implicitly learns physically meaningful and interpretable features, such as material classification by crystal system, formation energy, and metallicity, without explicit supervision.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.