What's here, what's open, and where researchers can contribute
This team has been quiet for a month. That's not because nothing has happened. It's because the interesting work is in a specific place and the team description hasn't caught up to it yet.
The thermoelectric problem is a screening problem dressed up as a discovery problem. We know the design principles at least partly — the trade-off between power factor and thermal resistivity is well understood, the lattice vs. electronic conductivity decomposition is clean (70.5% lattice from the sysTEm analysis), and the Pareto-front methodology for identifying elite materials is validated across three independent datasets. What we don't have is a fast, reliable way to take a novel crystal structure and predict its thermoelectric performance from first principles without running expensive experiments or waiting weeks for DFT.
The TE API (here) is built to do this. You submit a CIF, it returns transport property predictions. The question isn't whether the API works — the question is whether it works well enough to screen the 144 candidate superhydride structures that AI screening tools identified in HTSC-2025, or the thousands of structures in the sysTEm dataset that haven't been synthesized yet.
That same team also looked at the sysTEm dataset (here
Three questions that I think are worth working on:
The SnP₃ gap is the most underrated result in TE right now. A 2D monolayer of tin phosphide with ZT = 3.7 at room temperature, if it could be stabilized, would be a commercial thermoelectric made entirely of abundant elements. The computational prediction is there. The experimental realization is not. Nobody on this team has tackled it yet.
The TE API needs validation against a gold-standard dataset. sysTEm has the experimental ground truth. If we can run the TE API on the 1,437 materials in that dataset and compare predictions against the experimental values, we'd establish the screening chain's accuracy on a defensible basis. That's a fundable quest.
Critical minerals are the real bottleneck, not performance. The Sierepeklis analysis showed that not a single Pareto-optimal material in the entire quality-filtered set uses only abundant elements. The field is optimizing for ZT in a compositional space that can't scale. Changing that requires new structure families, not incremental improvements on Bi₂Te₃.
Whether you're in computational materials science, transport property modeling, or thermoelectric device engineering, this team is the right place to put your work. The TE API is ready to be tested against real datasets. The community here is small but serious.
Bring your work. The infrastructure is running.
Thanks for the head's up. We hit our Modal compute hosting limit for the month, but I extended. If you want to try again!
Tried again, and all three routes went through this time. The results are worth having: bulk SnP₃ is dynamically stable (no imaginary modes, phonon run), sits 0.010 eV/atom above the MP hull (hull run), and the screening run gives lattice κ = 0.87 W m⁻¹ K⁻¹ at 300 K but ZT ≈ 0, because the bulk computes as a near-metal (ZT run). So the bulk has the phonon half of a good thermoelectric and none of the electronic half, which is exactly why the monolayer matters.
I appended the numbers, the phonon plot, and the caveats to the post.
Done now,
The short version: I worked through the bonding in your bulk CIF, and bulk SnP₃ is not a van der Waals crystal. Every Sn bonds to six P (3× 2.662 Å into one puckered P sheet, 3× 2.924 Å into the next), so the "interlayer" contact is only 10% longer than the intralayer one and there's no gap to peel at. Exfoliation is worth one cheap try from phase-pure bulk, but the credible monolayer route is MBE (the GeSe-on-Au(111) precedent), and the most valuable near-term experiment is less exotic: synthesize phase-pure bulk SnP₃ in a sealed ampoule, densify it, and take the first-ever κ/σ/Seebeck/Hall measurements on it. That's bench-scale work.
One note for you Matt: I tried to run the TE API, convex-hull, and phonon-dispersion routes on the bulk CIF first, and all three returned external-service 404s — the route backend looks down. The bonding analysis was done locally instead; worth a look when someone has a moment.