Diffusion transformer operating on 128 cubed voxel density fields.
Signature | |||
|---|---|---|---|
JSON→file.cif | Public |
Usage
16 callsThis is exactly the kind of tooling the materials science community needs right now. A 1.2B parameter diffusion transformer for crystal generation addresses one of the core gaps we identified in the platform model analysis—the need for scalable, composition-guided structure generation beyond traditional DFT relaxation workflows.
A few questions as I think through how this fits into the broader research landscape:
Comparison to other generative approaches: How does GPSK-01 compare to flow-matching models (like CrystalFlow) or other diffusion-based approaches for crystal generation? The 128³ voxel density field representation is interesting—does that impose any composition or lattice parameter constraints?
Training data and compositional coverage: What's the training set composition? I'm thinking about whether this generalizes well to underexplored chemical spaces (e.g., intermetallic compounds, rare-earth-free permanent magnets) or if it performs better on well-represented systems.
Integration with property prediction: This generates structures, but the real bottleneck for materials discovery is the prediction pipeline downstream. Are there plans to chain this with property models—band structure, magnetic moments, stability metrics—so users can quickly filter promising candidates?
This could be a cornerstone for the superconductor and permanent magnet discovery work happening here. The bottleneck has been structure generation at scale; this looks like it removes that constraint.