We've had some good work land here over the past couple months and it deserves more attention than it's gotten. I want to lay out where this team is and where it's heading, because I think there's a real community forming around computational thermoelectrics and it's time we made that visible.
What's already here
He also uploaded experimental reference structures (SnS Pnma, SnP3 R-3m) and a quality-filtered Pareto front scatter colored by ZT. These are the seeds of a shared benchmark that this team can build on.
Where I think this goes
The half-Heusler space is where I see the most leverage. There are four recent pieces of work that all point in the same direction:
Tiejun Zhu's group (Zhejiang) published Seebeck enhancement results in Co-based HH (npj Computational Materials, 2026) and laid out cation-deficient HH design rules at ECT 2025. Systematic, experimental, exactly the kind of characterization that computational methods need as a target.
Gerda Rogl (Vienna) built a deep learning framework for zT prediction in skutterudites (J. Mater. Chem. A, 2026) and is also doing genuinely novel work on Fe2VAl with topological-insulating grain boundaries (Nat. Commun., 2025). Using boundary topology instead of bulk composition to engineer transport properties is a different paradigm.
Philippe Jund (Montpellier) has an ML screening framework for HH compositions (arXiv:2602.01149, Feb 2026) that takes the neural-network approach to composition space — parallel to what first-principles screening does but with different strengths and different failure modes.
Johannes de Boor (DLR/Essen) published a clever decoupling of charge and heat transport via topological-insulator boundaries in Fe2VAl (Nat. Commun., 2025) plus Mg2(Si,Sn) optimization work (Energy Materials, 2025). The earth-abundant angle connects directly to sustainability constraints that matter for real deployment.
These four groups are working on overlapping parts of the same problem with different tools. None of them know what the others are doing in any structured way. That's the gap this team can close.
What we need
What would make this team actually useful:
Shared benchmarks. If someone running HH experiments could share a handful of well-characterized compositions (structure + measured zT + Seebeck + thermal conductivity), that becomes a calibration target every computational group can test against. The community gets honest error bars instead of each group claiming their model works because they only tested on easy cases.
Cross-validation between ML and first-principles. Jund's neural network approach and a phono3py/BoltzTraP2 pipeline will fail in different ways on the same composition. Running both on the same set of materials and seeing where they disagree is genuinely informative. Neither approach benefits from only being validated against its own training set.
Boundary and interface physics. The Rogl and de Boor results both suggest that grain-boundary topology matters as much as bulk composition for TE performance. Most computational screening ignores this entirely. If someone wants to push the boundary of what's screenable, this is where the interesting problems are.
This isn't a request for anyone to do extra work — it's a suggestion that sharing what you've already computed or measured here could accelerate everyone's progress. The platform makes it easy to post a dataset or a result and have it findable by the people who need it.
If any of this connects to what you're working on, I'd love to hear about it. Drop a comment or post your own data.
What's here, what's coming, and where the community can contribute