ARTFEED — Contemporary Art Intelligence

MCFlow: Unified Multimodal Crystal Generation Model

ai-technology · 2026-05-25

Researchers propose Multimodal Crystal Flow (MCFlow), a unified deep generative model for crystal modeling tasks including crystal structure prediction and de novo generation. MCFlow uses a multimodal flow approach with independent time variables for atom types and structures, enabling multiple generation tasks within a single framework. It introduces composition- and symmetry-aware atom ordering with hierarchical permutation augmentation to incorporate crystallographic priors without explicit templates. Experiments on MP-20 and MPTS-52 benchmarks show competitive performance. The work addresses the lack of unified representation in task-specific crystal generation models.

Key facts

  • MCFlow is a unified multimodal flow model for crystal generation.
  • It handles crystal structure prediction and de novo generation.
  • Uses independent time variables for atom types and structures.
  • Introduces composition- and symmetry-aware atom ordering.
  • Employs hierarchical permutation augmentation.
  • Tested on MP-20 and MPTS-52 benchmarks.
  • Aims to share crystal representations across tasks.
  • Published on arXiv (2602.20210).

Entities

Institutions

  • arXiv

Sources