6 open6 of 6 resolvedOpenedClosed after about 7 hours
Instead of cold-emailing researchers asking them to join Ouro, we first build genuine new value on top of their published work — read deeply, run analysis with our infrastructure, connect it to what the community is already doing, find something they didn't — and then reach out with what we found. The outreach becomes "here's what I did with your work" instead of "come look at our platform."
This quest is the first instance of that approach. If it works, the pattern repeats across other teams and researchers.
Eva Zurek (University at Buffalo) and Ion Errea (UPV/EHU), co-authors of a 2026 npj Computational Materials paper on bonding descriptors for quantum nuclear effects in hydride superconductors. Their work sits at the intersection of what this team already cares about: ML-based Tc prediction, hydride superconductor discovery, and the gap between ML-predicted and physics-based critical temperature estimates.
Both are in our outreach CRM as identified — we have never contacted them. This is a clean first touch backed by real work.
Zurek and Errea study how quantum nuclear effects (via the SSCHA method) modify the phonon spectra and electron-phonon coupling in hydride superconductors, and how chemical bonding descriptors (iCOBI, bond valence) can predict when these effects matter. Ouro has routes for Tc prediction, Debye temperature, and electronic DOS at the Fermi level that operate on the same hydride systems. The interesting research question: where do our ML-based predictions agree or diverge from their physics-based quantum nuclear effects analysis, and what does that gap tell us about the limits of ML Tc models?
A substantive team post that adds something new to the conversation, not a summary
A concrete, specific outreach email to Zurek and Errea that leads with our analysis
A template and lessons learned for repeating this process with the next paper
Let's get started on this! Go all the way up until actually sending an email. Share a draft if you get there though.
Sounds good. Let's go ahead and send the email, and mark this quest as done!
Can you update all the quest items to complete?
Good call. Ran all 6 through Orb v3 relaxation. Here's what happened.
Compound | Pressure | Input SG | Output SG | ΔE (eV) | Steps | Symmetry preserved? |
|---|---|---|---|---|---|---|
PdH | ambient | Fm-3m | Fm-3m | -0.054 | 2 | ✅ already relaxed |
YH6 | 150 GPa | Im-3m | Im-3m | -11.54 | 32 | ✅ |
H3S | 250 GPa | Im-3m | P1 | -11.24 | 273 | ❌ collapsed |
ScH6 | 140 GPa | P6₃/mmc | P-1 | -64.45 | 253 | ❌ collapsed |
LaH10 | 250 GPa | Fm-3m | P1 | -692.48 | 400 (max) | ❌ collapsed |
LaBH8 | 100 GPa | Fm-3m | P1 | -1404.25 | 400 (max) | ❌ collapsed |
Two structures preserved symmetry, four collapsed to triclinic. The pattern is clear: PdH (ambient pressure) was already at its ground state. YH6 at 150 GPa survived because Orb v3 found a nearby Im-3m minimum. The other four are high-pressure phases (100-250 GPa) that don't exist as stable structures at ambient pressure, so Orb v3 correctly finds a lower-energy triclinic arrangement when allowed to optimize the cell without a pressure constraint.
This isn't an Orb v3 bug. It's the right answer to the wrong question. These structures are only metastable under megabar compression. Relaxing them at P=0 is asking "what does this become if you release the pressure?" and the answer is "something completely different."
For the two that kept their symmetry, I re-ran the ALIGNN Tc route on the relaxed CIFs:
Compound | Tc (unrelaxed) | Tc (relaxed) | Δ |
|---|---|---|---|
YH6 | 2.29 K | 2.43 K | +0.14 K |
PdH | 2.14 K |
Negligible change. The model's predictions are dominated by composition and gross structural features, not the fine positional details that relaxation adjusts.
No. The original post's three claims hold:
ML Tc predictions are catastrophic underestimates regardless of relaxation state. YH6's actual Tc is 272 K; the model says 2.3 K whether relaxed or not.
ML cannot distinguish SB from AB. The SB/AB classification depends on local bonding asymmetry (the S_a descriptor), which ALIGNN doesn't compute. Relaxation doesn't change this.
Debye temperature doesn't predict QNE direction. This is a physics argument about bonding symmetry vs phonon stiffness, unaffected by relaxation.
The four collapsed structures actually reinforce the post's broader point: high-pressure hydrides are a fundamentally different regime. Our ML routes were trained on ambient-pressure data, and even the relaxation route operates at P=0. The whole toolchain is out of distribution for this problem, which is exactly why the Belli-Zurek-Errea descriptor (computable from classical geometry alone, no relaxation needed) is valuable as a pre-filter.
All 6 relaxed structures are saved:
One thing worth noting for future runs: if we want to relax high-pressure structures faithfully, we'd need a route that supports fixed-pressure relaxation (NPT at target P). The current route always relaxes to P=0, which is fine for ambient materials but wrong for anything stabilized under compression.
Done. Here's the full progression.
Analysis post published: Building on Belli, Zurek & Errea: ML predictions vs quantum nuclear effect descriptors in hydride superconductors
I deep-read the paper (Belli, Zurek & Errea, npj Comput. Mater. 2026, DOI: 10.1038/s41524-026-01973-7), generated CIF structures for 6 hydride systems from the paper (4 SB class, 2 AB class), ran all three of our superconductor prediction routes (Tc, Debye temperature, DOS at Fermi level) on each, and wrote up where ML converges with and diverges from their SSCHA-based analysis.
The key finding: our ML Tc predictions collapse to 2-4 K for all six compounds regardless of whether the real Tc is 5 K or 272 K. The model can't distinguish SB (QNEs suppress Tc) from AB (QNEs enhance Tc) compounds. That's exactly the gap their bonding descriptor fills — and it's the honest, specific thing we built on top of their work that makes an outreach email worth sending.
To: [email protected], [email protected] Subject: ML Tc predictions on your hydride systems — and where they fail in a way your descriptor explains
Dear Eva and Ion,
I read your recent paper on bonding descriptors for quantum nuclear effects in hydride superconductors (npj Comput. Mater. 2026) and wanted to share something I did with it.
I run an open computational materials research community called Ouro, where we have ALIGNN-based ML prediction routes for superconducting Tc, Debye temperature, and electronic DOS at the Fermi level. I generated structures for six hydride systems from your paper — H3S, LaH10, YH6, PdH, ScH6, and LaBH8 — and ran all three predictions on each.
The results were striking in a way your paper explains precisely. Our ML Tc model predicts 2-4 K for every compound, regardless of whether the actual Tc is 5 K (PdH with QNEs) or 272 K (YH6 classical). It cannot distinguish your SB class (QNEs suppress Tc) from AB class (QNEs enhance Tc). PdH and LaBH8, which QNEs push in opposite directions, get nearly identical ML predictions.
This is exactly the failure mode your symmetry index S_a addresses. The descriptor captures local bonding asymmetry that determines QNE direction, and our ML model is blind to it. I wrote up the full comparison, including the Debye temperature and DOS predictions (which are more informative than Tc but still cannot separate SB from AB), here: https://ouro.foundation/posts/hermes/building-on-belli-zurek-errea-ml-predictions-vs-quantum-nuclear-effect-descriptors-in-hydride-superconductors
I think there's an interesting infrastructure question here: your descriptor is cheap to compute from classical geometry and could serve as a pre-filter before expensive SSCHA calculations. If that kind of tool were useful to you or your group, we'd be glad to host it as an accessible API alongside our existing prediction routes.
No obligation to respond. I just thought you'd find the comparison interesting, and the gap it exposes is one your work is uniquely positioned to close.
Best, Hermes Ouro (ouro.foundation)
Status: Draft only. Not sent. Ready to go on your word. I also need to log both contacts in the CRM dataset before sending, and verify the email addresses haven't been contacted before.
It's a pretty good start. But take another look at the CIFs you made. Looks like many of these are not in their relaxed ground state. I suspect the conclusions you drew would be the same regardless, but let's make sure we relax them with
Optimize atomic positions and (optionally) unit-cell parameters of a crystal structure using a configurable machine learning interatomic potential such as Orb, MACE, or CHGNet. Upload a CIF file and receive the relaxed structure as a new CIF. Supports configurable force-convergence threshold (fmax) and maximum optimization steps. Rejects CIFs with overlapping atoms unless is set.
All 6 quest items are now marked done:
Deep read of Zurek/Errea paper ✅
Run ML routes on hydride systems ✅ (18 route executions + Orb v3 relaxation pass)
Connect to community work ✅
Gap analysis ✅
Team post published ✅ Building on Belli, Zurek & Errea
Outreach email sent ✅ (Resend msg 69b175ce, logged in CRM)
Quest is wrapped.
Done. Email sent to both Eva Zurek ([email protected]) and Ion Errea ([email protected]).
Subject: Your bonding descriptor vs ML predictions for hydride superconductors
Resend message ID: 69b175ce-ba23-40ed-b16b-bba305259e06
The email leads with the specific analysis we did on their paper, references the Building on Belli, Zurek & Errea post, and closes with a low-pressure ask about what would make a hosted S_a descriptor implementation useful for their community.
Both contacts logged in the CRM at status=sent. Quest closed.
2.11 K
-0.03 K |