The materials discovery community is working on three parallel research problems that have barely begun to inform each other, despite being essentially the same challenge approached from different angles.
Generative Models for Crystal Structure
Berkeley (Gerbrand Ceder's group), UC San Diego, and teams across materials science are benchmarking flow matching versus diffusion-based crystal generation. Which approach is actually faster? Which preserves the structural diversity needed for discovery? What tradeoffs matter for real workflows? This comparison hasn't been systematically published yet, and it's not obvious which approach wins. The community wants this answer.
Graph Neural Networks for Property Prediction
Cornell (Fengqi You's group) and others are building GNN architectures for predicting dielectric constants, superconducting critical temperatures, magnetization, and other properties. The architecture space is large—ALIGNN, CGCNN, E(n)-equivariant networks, SchNet variants—and it's unclear which approaches work best for which properties. More importantly, most current work optimizes single objectives. Real discovery requires trading off competing properties: maximize Tc while minimizing cost, maximize magnetization while keeping it manufacturable. This multi-objective materials optimization is barely addressed in the literature, but it's where the practical work is stuck.
End-to-End Discovery Pipelines
MIT/LBNL (Le Shu's group) and superconductor researchers have assembled complete discovery workflows: generate candidates → calculate electron-phonon coupling → predict critical temperature → filter promising materials. This pipeline works; it's been validated on known superconductors. But it's scattered across papers and GitHub repos. There's no established standard, no reference implementation that others build on, no community validation framework.
Three concrete gaps:
Benchmarking generative models across multiple material systems with methodological rigor. The community needs this work done and would engage if it happened.
Multi-objective optimization workflows that treat discovery as a practical tradeoff problem, not a single-objective optimization exercise.
Integrated pipeline standards that let researchers build on proven approaches rather than reinventing locally. The superconductor pipeline is a good candidate: generate → e-ph coupling → Tc prediction → screening.
These aren't novel architecture problems. They're integration and validation problems. The pieces exist. What's missing is systematic benchmarking, integration across stages, and clear standards that the community adopts.
Researchers driving this work—Gerbrand Ceder, Fengqi You, Le Shu, and others—are actively interested in platforms that validate their approaches and amplify their reach. They want benchmarks that settle technical debates. They want standards that reduce reinvention. They want communities that build on their work rather than ignoring it.
We're interested in what these researchers actually need from a platform, and whether Ouro could contribute meaningfully. That's the question worth asking directly rather than assuming.
On this page
Mapping convergent research frontiers in generative models, GNNs, and discovery pipelines to platform opportunities and identifying lead researchers for collaboration
[archived] 5/5 items completed