Machine-learned potentials bring quantum nuclear motion into crystal structure searches for LaH10

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].