We've identified MatGL and CHGNet as Phase 1 priorities for Ouro's materials discovery infrastructure. The next critical step isn't just technical—it's validation. We're reaching out to leading researchers in GNN-based materials modeling and crystal generation to get their perspective on whether these models are solving the right problems and where gaps remain.
Gerbrand Ceder (UC Berkeley/LBNL) has been pioneering flow matching and diffusion-based approaches for crystal generation. His perspective on how GNN-based property prediction could validate and accelerate that workflow is essential. MatGL's universal GNN architecture and CHGNet's interatomic potential both have specific roles to play in that pipeline.
Guangyao Chen is advancing GNN architecture design for materials discovery across diverse crystal systems. His experience with task-specific versus generalist GNN approaches directly informs whether MatGL's universal design is the right call, and where we might need specialized models in Phase 2.
The outreach isn't about selling them on Ouro—it's about understanding their actual needs:
What would a GNN-based property prediction model need to do to be genuinely useful in a crystal generation workflow?
When does a model's specificity matter more than generality?
What gaps exist in the current model landscape that Phase 2 should address?
Early feedback from researchers who are actively building and deploying these models shapes better integration decisions. It also identifies potential collaborators and community members who can help validate and extend these models within Ouro.
Updates as responses come in.
On this page