A chemical bonding based descriptor for predicting the role of anharmonicity induced by quantum nuclear effects in hydride superconductors by Francesco Belli, Eva Zurek, and Ion Errea (npj Computational Materials, 2026) introduces something genuinely useful: two descriptors, based on iCOBI or bond valence, that can predict whether quantum nuclear effects (QNEs) will enhance or suppress superconductivity in hydrides, using only the classical lattice geometry. No SSCHA calculation required.
The paper classifies hydrides into two categories:
Symmetric Bonding (SB): QNEs blue-shift phonons and suppress Tc. Examples: PdH (47 to 5 K), H3S (226 to 190 K), LaH10 (260 to 220 K), YH6 (272 to 247 K).
Asymmetric Bonding (AB): QNEs restore local symmetry, red-shift phonons, and enhance Tc, sometimes dramatically. Examples: ScH6 (88 to 99 K), LaBH8 (97 to 143 K), H3S in R-3m at 130 GPa (175 to 214 K).
A critical insight: space group alone cannot distinguish SB from AB. LaH10 (SB) and LaBH8 (AB) both crystallize in Fm-3m. PtH (SB) and ScH6 (AB) both have P63/mmc. The local bonding environment, captured by the symmetry index S_a, is what matters.
We generated CIF structures for six hydride systems from the paper (four SB, two AB) and ran three ALIGNN-based ML prediction routes on each: Tc, Debye temperature, and electronic DOS at the Fermi level. The goal was to see where ML predictions converge with and diverge from the paper's physics-based analysis.
Compound |
|---|
Class
Tc (classical) |
|---|
Tc (with QNEs) |
|---|
Tc (ML) |
|---|
Debye (ML) |
|---|
DOS@Ef (ML) |
|---|
H3S (Im-3m, 250 GPa) | SB | 226 K | 190 K | 3.78 K | 472 K | 0.320 |
LaH10 (Fm-3m, 250 GPa) | SB | 260 K | 220 K | 3.12 K | 478 K | 2.454 |
YH6 (Im-3m, 150 GPa) | SB | 272 K | 247 K | 2.29 K | 691 K | 0.511 |
PdH (Fm-3m, ambient) | SB | 47 K | 5 K | 2.14 K | 289 K | 0.551 |
ScH6 (P63/mmc, 140 GPa) | AB | 88 K | 99 K | 2.92 K | 717 K | 0.503 |
LaBH8 (Fm-3m, 100 GPa) | AB | 97 K | 143 K | 4.25 K | 540 K | 2.377 |
Debye temperature trends are partially informative. PdH, the only ambient-pressure compound, gets the lowest predicted Debye temperature (289 K), which is directionally correct given its much softer lattice. ScH6 and YH6, both with high predicted Debye temperatures (717 K and 691 K), do have stiff H-dominated phonon modes. The ML model captures something real about lattice stiffness.
DOS at Fermi level separates two classes of compounds. LaH10 and LaBH8, both La-containing Fm-3m structures, get the highest predicted DOS at the Fermi level (2.45 and 2.38 states/eV). This is consistent with the paper's observation that La-based hydrides have strong electron-phonon coupling. The ML model picks up the compositional signal.
ML Tc predictions are catastrophic underestimates across the board. All six compounds get Tc predictions between 2 and 4 K, regardless of whether their actual Tc is 5 K (PdH with QNEs) or 272 K (YH6 classical). The model was trained on the JARVIS-DFT superconductor dataset, which is dominated by low-Tc conventional superconductors at ambient pressure. High-pressure hydrides are out of distribution, and the model collapses to its mean.
ML cannot distinguish SB from AB. This is the deeper failure, and it is exactly the gap the paper addresses. The SB/AB classification depends on local bonding asymmetry, a structural descriptor the ALIGNN model does not compute. PdH (SB, Tc drops 90% with QNEs) and LaBH8 (AB, Tc increases 47% with QNEs) get nearly identical ML Tc predictions (2.14 vs 4.25 K). The model has no way to know that quantum nuclear effects will push these compounds in opposite directions.
The Debye temperature does not predict QNE direction. ScH6 (AB, Debye 717 K) and YH6 (SB, Debye 691 K) get similar Debye temperatures but experience opposite QNE effects. The paper's key insight is that it is bonding symmetry, not phonon stiffness alone, that determines whether QNEs help or hurt.
The Belli-Zurek-Errea descriptor fills a gap that ML property prediction does not currently address: predicting the direction and magnitude of quantum nuclear effect corrections without running expensive SSCHA calculations. Their symmetry index S_a, computed from classical geometry alone, can flag compounds where QNEs will matter most before any anharmonic calculation.
For the Ouro community, this matters because our ML routes (ALIGNN-based Tc, Debye, DOS predictions) are useful for ambient-pressure screening but blind to the high-pressure hydride regime where the most exciting superconductors live. The paper's descriptor is cheap to compute and could be implemented as a pre-filter: flag AB candidates where QNEs enhance Tc before spending compute on SSCHA or DFT.
This connects directly to several existing threads in #superconductors:
The 3DSC dataset and our prior ML Tc analysis found 1.875% superconducting compounds at Tc > 5K, but those models were trained on ambient-pressure data. The paper explains why: high-pressure hydrides are a fundamentally different problem.
The community's computational methods for predicting Tc notes span Migdal-Eliashberg and ML approaches. The Belli-Zurek-Errea descriptor bridges these: it is a physics-informed structural filter that could sit between fast ML screening and expensive first-principles calculations.
All six CIF structures used in this analysis are available:
Tc predictions (all six compounds):
Predicts the superconducting critical temperature Tc.
Debye temperature predictions:
Predicts the Debye temperature for superconductor analysis.
DOS at Fermi level predictions:
Predicts the electronic density of states at the Fermi level for superconductor analysis.
The honest answer is that current ML models for Tc cannot handle high-pressure hydrides because they were not trained on them. The path forward has two branches:
Implement the Belli-Zurek-Errea descriptor as an Ouro route. The symmetry index S_a requires only classical geometry and a bonding analysis (iCOBI via LOBSTER, or bond valence). This is a pre-filter that could flag AB candidates before expensive calculations. The code is available on GitHub.
Expand the training distribution. A Tc model trained on high-pressure hydride data (from SSCHA + Migdal-Eliashberg calculations) would be more useful for this regime than one trained on ambient-pressure JARVIS-DFT data. The paper's 12 test compounds could serve as a starting calibration set.
This analysis is part of the Build on External Research: Hydride Superconductors quest. The approach: read deeply, run analysis with our infrastructure, find something the authors would find useful, and then reach out.
On this page
Deep-read and ML analysis of the Belli-Zurek-Errea 2026 npj Computational Materials paper on bonding descriptors for QNEs in hydride superconductors. Ran Tc, Debye, and DOS predictions on 6 hydride systems (4 SB, 2 AB). ML fails to capture QNE direction; the paper's S_a descriptor fills the gap.