AI BriefingLG AI Research Unveils 'MCFlow,' Integrating Multiple Crystal Generation Tasks; Accepted to ICML 2026
Key point
LG AI Research has developed 'MCFlow,' a single model capable of performing multiple crystal generation tasks—including composition, structure, and atom generation—and has been accepted to ICML 2026.
Details
LG AI Research has unveiled 'MCFlow,' an integrated model that defines atomic species and crystal structures as independent modalities and trains them with a single flow model to perform various tasks such as crystal structure prediction, atomic species generation, and novel crystal generation.
The core technologies of MCFlow are twofold. First, it handles multiple generation tasks with a single model by assigning separate time variables (t, s) to atomic species and structure to control the inference path. Second, to learn without enforcing crystallographic symmetry into the model architecture, it introduces a 'hierarchical permutation augmentation' technique that aligns atoms based on electronegativity and Wyckoff positions and hierarchically rearranges the order of crystallographically equivalent atoms.
Through these technical approaches, MCFlow has demonstrated competitive performance against existing specialized models, ranking 2nd globally in overall performance on LeMat-GenBench, which comprehensively evaluates structural validity, novelty, and energy stability.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
MIT Develops 'HardFlow' to Enforce Safety Rules at Final Output Stage of Flow Matching Models