Science1 publisherNot yet confirmed elsewhere2 min readPublished
Machine-learned potentials bring quantum nuclear motion into crystal structure searches for LaH10
Poletaev and Oganov tested three machine-learned potentials inside SSCHA-based evolutionary structure searches, using the superconducting hydride LaH10. Their comparison gives hydride groups a first guide to trading training cost against accuracy, drawn from one compound whose answer was already known.
The Scientist · Science desk

What happened
- Active-learning potentials trained on the fly recovered LaH10's cubic Fm-3m phase but needed thermodynamic perturbation theory corrections to give consistent results.
- The foundation model Mattersim-5m ran SSCHA-based searches without per-structure training, though the authors say fine-tuning is required for higher accuracy.
- Including quantum anharmonicity simplified LaH10's free-energy landscape and was essential for putting its candidate structures in the correct energy order.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- cost The active-learning route adds work at both ends of a search: the potential is trained during the run, and the results then need a perturbation-theory correction.
- decision A group with no SSCHA data of its own can start searching with Mattersim-5m, but has to budget for fine-tuning before relying on its energy rankings.
- exposure Hydride structure rankings computed without quantum anharmonicity are open to revision, if the LaH10 result carries over to other compounds.
The ranking finding is why the software comparison matters. If quantum nuclear motion decides which candidate sits lowest in free energy, a search that leaves it out can put the wrong structure on top [9]. Poletaev and Oganov argue that this matters most for light-atom systems such as superconducting hydrides, and that reliable finite-temperature prediction with quantum anharmonic effects included has stayed hard [4]. Their fix is to put machine-learned potentials inside SSCHA, so an evolutionary search runs directly on the quantum anharmonic free-energy landscape [3].
The three potentials differ mainly in when the training happens. Active-learning potentials are trained on the fly, inside the search [5]. Mattersim-5m is a foundation model and needs no training for each new structure [7]. The temperature-dependent effective potentials are trained on SSCHA ensemble data [5], so those ensembles have to be generated before the potential exists. Only after that step does it deliver its speed on large unit cells [8]. In a screening budget, training is a one-off cost and the search is a per-candidate cost, and the three routes split the two very differently. Two of the three also needed an extra step, corrections or fine-tuning, before their output was consistent or more accurate [12].
The thing this doesn't tell you is how the methods behave when the answer is unknown. LaH10 is a test with a target, and the cubic Fm-3m phase is what "correct" means here [6]. Recovering a known structure is a recall test. A screening campaign asks a method to rank many unfamiliar compositions, and one compound cannot show how often each route gets that right. The abstract does not report run times or error figures, so "very efficient" cannot yet be converted into core-hours per candidate [8].
I think the temperature-dependent route is the one to back for large cells, on two conditions. The group needs SSCHA ensembles for the chemistry it is screening, or the budget to make them [5]. And the LaH10 result has to repeat on a hydride whose structure was not known in advance, since every comparison in the paper used that single compound [5][8].
The Russian Science Foundation funded the work under grant 19-72-30043 [1]. Calculations ran on the Oleg and Zhores clusters at the Skolkovo Institute of Science and Technology, with some on shared HPC facilities at UNN [10]. The authors declare no competing interests [2], and the paper is open access under a Creative Commons Attribution 4.0 licence [11].
What to watch
- A run of the SSCHA-trained temperature-dependent potential on a hydride whose stable structure is not already known, with the prediction later checked by experiment or higher-level calculation.
- Published timings for the three routes, including the cost of generating the SSCHA ensembles, which would let groups price a screening campaign.
- Whether a fine-tuned Mattersim-5m reaches the same LaH10 ranking as the corrected active-learning potentials.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption
- Insufficient
- Hype gap+10
- Incentives15
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The work was supported by the Russian Science Foundation, Grant No. 19-72-30043.
ReportedSupportedSource: Paper funding statement2 sources— create a free account to open themView cited source - [2]
The authors declare no competing interests.
ReportedSupportedSource: Paper ethics declarations2 sources— create a free account to open themView cited source - [3]
Poletaev and Oganov, in npj Computational Materials (2026), integrate machine-learned interatomic potentials (MLIPs) with the stochastic self-consistent harmonic approximation (SSCHA) to enable evolutionary crystal structure prediction on the quantum anharmonic free-energy landscape.
- [4]
Reliable crystal structure prediction at finite temperatures, including quantum anharmonic effects, remains challenging but is particularly important for systems containing light atoms such as superconducting hydrides.
- [5]
Using LaH10 as a test case, the authors compare three approaches: active-learning MLIPs trained on the fly, universal MLIPs, and temperature-dependent effective potentials (TDEPs) trained on SSCHA ensemble data.
- [6]
Active-learning MLIPs correctly predict the cubic Fm-3m phase of LaH10 but require thermodynamic perturbation theory corrections for consistent results.
- [7]
The foundation universal MLIP Mattersim-5m enables SSCHA-based crystal structure prediction without per-structure training, but fine-tuning is required for higher accuracy.
- [8]
Temperature-dependent effective potentials trained on SSCHA data enable accurate and very efficient crystal structure prediction for large unit cells on the quantum anharmonic free-energy landscape.
- [9]
"We demonstrate that quantum anharmonicity simplifies the free-energy landscape and is essential for the correct energy ranking of LaH10 structures."
- [10]
Calculations with MTPs were performed on the Oleg cluster at the Skolkovo Institute of Science and Technology; calculations with uMLIPs including Mattersim-5m, and training and using TDEP-DPA3, were performed mainly on the Zhores cluster at Skoltech and partly on HPC SEC shared research facilities at UNN.
- [11]
The article is open access under a Creative Commons Attribution 4.0 International License.
- [12]
Two of the three approaches tested needed an additional step: perturbation theory corrections for the active-learning potentials and fine-tuning for Mattersim-5m.
Sources
1 independent publisher whose own reporting we read for this story.
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.