LG AI Research Develops Integrated AI Model 'MCFlow' for Inorganic Material Generation; Accepted to ICML 2026

LG AI Research has developed MCFlow, an integrated model for inorganic material generation that defines atomic species and crystal structures as independent modalities, and it has been accepted to ICML 2026. Unlike existing models specialized for single tasks, MCFlow can perform three tasks with a single model: crystal structure prediction, novel generation, and atomic species generation.

Key Technologies and Features

MCFlow integrates atomic species (Atom Type) and structure (Structure) by separating them into different modalities and assigning independent generation time axes. This enables Atom Substitution, where specific elements can be changed while maintaining the existing structure. This feature can be directly applied to material improvements at the commercialization stage, such as transition metal substitution in battery cathodes or modification of catalytic active sites.

Additionally, the research team proposed a Hierarchical Permutation Augmentation strategy that arranges atoms based on electronegativity and Wyckoff Position while considering crystallographic equivalence, allowing the model to effectively learn crystal symmetry without enforcing space groups. Through these technical approaches, MCFlow achieved a world ranking of 2nd overall on the LeMat benchmark, which comprehensively evaluates structural validity, novelty, uniqueness, and energy stability.

Future Plans

LG AI Research plans to improve a demo page that allows users to select generation tasks and set conditions via natural language input. In the long term, the goal is to add Property as a new modality and expand the scope of integrated generative AI beyond inorganic materials to organic materials (molecules, polymers, etc.).