Materials

JANUS: A Multi-modal Foundation Neural Sampler for Disordered Materials

Denis Blessing, Mouyang Cheng, et al.

Featured September 4, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new AI tool called JANUS learns to predict how messy materials behave by mixing and matching different atoms while also letting them wiggle and change volume, all without needing tons of pre-made data.

In depth

This paper introduces JANUS, a novel multimodal neural sampler that addresses the complex challenge of simulating disordered materials by jointly modeling discrete chemical identities and continuous structural fluctuations. It achieves this by coupling masked discrete diffusion for atomic species with continuous diffusion for atomic displacements and cell volume, all learned through an E(3)-equivariant graph neural network trained directly from energy evaluations of self-generated configurations.

Key Takeaways

1
JANUS unifies discrete chemical identities with continuous structural and volumetric relaxation, enabling joint thermodynamic sampling of disordered materials.
2
The model employs a self-bootstrapped training framework, learning directly from energy evaluations of its own generated configurations, eliminating the need for pre-generated equilibrium data.
3
It demonstrates amortized sampling across thermodynamic conditions and system sizes, significantly reducing computational cost for applications like phase diagram mapping, inverse design, and defect discovery.

Conceptual Flow

HIGH LEVEL

Methodology: How JANUS Works

JANUS learns to create new material structures by guessing atom types and positions at the same time, like filling in a puzzle while also letting the pieces move around.

Start with Masked Atoms

Start with Random Positions

Guess Atom Type & Adjust Position

New Material Structure

Results: What JANUS Achieves

JANUS can find new material recipes, predict how materials change with heat, and discover tiny flaws, all much faster than old methods.

Material Problem

Old Slow Way

JANUS Finds Solution Fast

New Materials

Phase Diagrams

Defect Designs

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.