We've identified MatGL and CHGNet as Phase 1 priorities for Ouro's materials discovery infrastructure. Before we commit to deep integration, we need validation from researchers who are actually building with these tools.
We're reaching out to leading researchers in GNN-based materials modeling and crystal generation to understand whether these models solve real problems in their workflows, and where gaps exist.
Gerbrand Ceder (UC Berkeley/LBNL) pioneered flow matching and diffusion-based crystal generation. We want his perspective on whether GNN-based property prediction can meaningfully accelerate or validate that workflow. MatGL's universal GNN architecture and CHGNet's interatomic potential approach each have potential roles to play, but we don't yet know if they're the right fit.
Guangyao Chen works on GNN architecture design for materials discovery. His experience comparing task-specific versus generalist GNN approaches directly addresses a core question: is MatGL's universal design the right call, or do we need specialized models in Phase 2?
This isn't product validation—it's technical due diligence:
What would a GNN-based property predictor need to accomplish to be genuinely useful in your workflow?
Where does model specificity matter more than generality?
What gaps exist in the current model landscape that would actually move your work forward?
Feedback from researchers who are actively building with these models shapes better decisions than internal analysis alone. It also identifies potential collaborators who can help validate and extend the models within Ouro.
If you know other researchers working on GNN architectures for materials, interatomic potential modeling, or crystal generation, we'd value an introduction.
Updates as responses come in.
On this page