Resolving the structural stability of the hydrogen storage borohydride NaCa(BH4)3 via data-driven screening and first-principles lattice dynamics
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: We're looking at the paper: "Resolving the structural stability of the hydrogen storage borohydride NaCa(BH4)3 via data-driven screening and first-principles lattice dynamics".
Mira: Here's the gist of the paper. Machine-learning screening and first-principles lattice dynamics resolve the most stable structure of NaCa(BH4)3 from a data-driven candidate search The most stable structure is a monoclinic Cm framework lying outside the perovskite family from which every earlier candidate for this composition was drawn Its ordered cation arrangement is the ground state of its own series, and lattice dynamics confirm that it is a minimum with respect to atomic displacement as well as to energy.
Kai: Discovery Strategy The authors assembled six candidate frameworks from three independent sources, including experimentally determined ABX3 borohydrides, a distorted perovskite from data-driven structure prediction, and an unbiased search over a structural database.
Mira: They enumerated all 420 decorations reducing to 118 symmetry-inequivalent configurations and screened them with a machine-learning potential before settling the most stable structure from first principles.
Kai: The search was controlled by fixing the charge-neutral composition and parameterizing cation arrangement by the swap fraction s, ranging from 0 to 1.
Paper summary: Kai: To wrap up on "Resolving the structural stability of the hydrogen storage borohydride NaCa(BH4)three via data-driven screening and first-principles lattice dynamics", they found that this compound has a most stable form that is a monoclinic framework, space group Cm, designated G2 > <ref:2609.02001#pg1,screening and first-principles lattice dynamics>
Mira: So what that means in simple terms is that even though we only have theoretical predictions for NaCa(BH4)three right now because it hasn't been made yet, the best prediction points to something completely different from what we were expecting based on previous structural guesses > <ref:2609.02001#pg1>
Lev: It solidifies the idea that the most stable arrangement of sodium and calcium ions in this borohydride isn't a simple perovskite structure, which is a lot for predicting new material stability >
Kai: And they’re testing it by saying that if you compare its energy to known synthesized compounds, specifically those around three point three seven meV per atom above the NaBH4 plus Ca(BH4)two tie-line, it falls right into the range of compounds that have actually been made > <ref:2609.02001#pg1>
Mira: So the implication is that this prediction isn't just some abstract theory; it’s a testable prediction, and it sits within the metastability range of compounds we already have in our labs >
Lev: The paper shows how machine learning can cover a library exhaustively, but you need those first-principles dynamics to decide which candidate is actually the ground state structure >
Kai: So for anyone interested in hydrogen storage materials, this study highlights that the structural search needs to go beyond just checking if a structure looks chemically intuitive; it has to check dynamic stability against atomic motion >
Conclusion: Kai: So, we're wrapping up on how they tackled the stability of NaCa(BH4)three using machine learning and then checking it with first principles lattice dynamics.
Mira: Yeah, they’ve done this whole process starting with a data-driven candidate search and ending up identifying a monoclinic structure outside the perovskite family.
Lev: It seems like the main point here is that machine learning is good for finding candidates but you need those first principles calculations to actually confirm if that candidate isn't just dynamically unstable.
Kai: Exactly, and what they found is that this most stable structure, designated G2, isn't a perovskite at all; it’s a monoclinic framework.
Mira: And the authors stress that this ordered arrangement of cations is the ground state for its own series of compounds. That's a big statement about how these materials organize themselves structurally.
Lev: From an error correction standpoint, if we had to run this on real hardware, we'd need to be really careful about those cation ordering details because they dictate the whole structure.
Kai: The implication is that the high-symmetry orthorhombic candidate they looked at earlier isn't just hard to make; it’s dynamically unstable.
Mira: And this new finding suggests that what they found here, this monoclinic framework, is actually stable in a way that fits right into the range of compounds we’ve already managed to synthesize.
Lev: The fact that it sits within the metastability range of existing compounds is pretty telling for any future synthesis plans.
Kai: So, basically, they're showing us a specific structure prediction that's not just theoretical fluff but something testable against what we can actually build in the lab.
Mira: It really highlights how important it is to couple those broad machine learning searches with the detailed physics of lattice dynamics for real material science.
Lev: And since they pointed out that this prediction is testable, it opens up a clear path for experimentalists to verify this structure using techniques like X-ray diffraction.
Sanaa Ismail, Ricardo Amaral, Attia A. Gadallah, Hassan Mohamed El-Said Azzazy, Zi-Kui Liu
Department of Chemistry, School of Sciences and Engineering, The American University in Cairo · Department of Materials Science and Engineering, The Pennsylvania State University
cond-mat.mtrl-sci
Submitted: 2026-09-02
Updated: 2026-10-30
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 70/100
The gist: Machine-learning screening and first-principles lattice dynamics resolve the most stable structure of NaCa(BH4)3 from a data-driven candidate search The most stable structure is a monoclinic Cm…
Key concepts
- Machine-learning screening
- The researchers used a machine learning potential to quickly screen 420 candidate structures. This process helped narrow down the vast number of possibilities to a manageable set of configurations before performing more rigorous, expensive first-principles calculations.
- First-principles lattice dynamics
- This method uses fundamental physical laws (like those in VASP) to analyze how atoms move within a crystal structure. The lattice dynamics confirmed that the found structure is dynamically stable, meaning it represents a minimum in both energy and atomic displacement.
- Perovskite family
- The most stable structure found was not a perovskite, despite earlier candidates being drawn from this family. This indicates that the most stable form of NaCa(BH4)3 belongs to a different structural class, challenging previous assumptions about its possible forms.
- Swap fraction (s)
- This parameter was used to systematically vary the cation arrangement in the search. By changing 's' from 0 to 1, the researchers explored how different types of Na and Ca ions could be ordered within the crystal lattice.
Terminology
Summary
Machine-learning screening and first-principles lattice dynamics resolve the most stable structure of NaCa(BH4)3 from a data-driven candidate search The most stable structure is a monoclinic Cm framework lying outside the perovskite family from which every earlier candidate for this composition was drawn Its ordered cation arrangement is the ground state of its own series, and lattice dynamics confirm that it is a minimum with respect to atomic displacement as well as to energy. p. 2 p. 0
Discovery Strategy
The authors assembled six candidate frameworks from three independent sources, including experimentally determined ABX3 borohydrides, a distorted perovskite from data-driven structure prediction, and an unbiased search over a structural database. They enumerated all 420 decorations reducing to 118 symmetry-inequivalent configurations and screened them with a machine-learning potential before settling the most stable structure from first principles. p. 0 The search was controlled by fixing the charge-neutral composition and parameterizing cation arrangement by the swap fraction s, ranging from 0 to 1. p. 2
Computational Methodology
The methodology involved a multi-stage process to ensure reliability, including candidate relaxation under a machine-learning protocol and subsequent re-relaxation from first principles. The cation ordering was resolved through two complementary enumerations, varying the supercell size and the swap fraction s. p. 9 Machine-learning energetics were used to rank configurations, but first-principles calculations with VASP were employed for final validation. p. 4 Lattice dynamics confirmed dynamical stability by showing that the unrestricted search returned a framework outside the perovskite family altogether, designated G2. p. 3
Structural Findings
The most stable structure identified is a monoclinic framework of space group Cm, designated G2. This structure was reached only by the unrestricted search, at −4. p. 0185640 eV/atom, 7. p. 067 meV/atom below the best framework available beforehand. p. 0 The ordered cation arrangement is the ground state of its own series. p. 0 Lattice dynamics complete the strategy, screened on the potential and settled from first principles. p. 3
Implications for Materials Science
The study demonstrates that machine learning is a right instrument for covering a candidate library exhaustively but the wrong one for deciding between its leading members. The findings suggest that the high-symmetry orthorhombic candidate is not merely difficult to prepare but dynamically unstable, whereas the structure identified here as most stable is not a perovskite at all. p. 3 The prediction is testable, and at 3. p. 0 p. 037 meV/atom above the NaBH4 + Ca(BH4)2 tie-line it sits well inside the metastability range of compounds that have in fact been synthesised.
Limitations and Future Directions
The study notes limitations, such as the near-zone-centre region where dispersion is neither computed directly nor constrained by the elastic tensor. The lattice dynamics are harmonic, so the anharmonic stabilisation invoked for the soft mode of B is cited from the literature rather than computed. p. 5 No decomposition pathway was computed, so nothing in this work bears directly on the dehydrogenation enthalpy or on the reaction products. p. 1 The structure of NaCa(BH4)3 remains a prediction, and the diffraction pattern given here is offered as the means of testing that prediction, not as evidence that it holds. p. 0
Conclusion
The most stable form of NaCa(BH4)3 is a monoclinic Cm framework lying outside the perovskite family from which every earlier candidate for this composition was drawn. Its ordered cation arrangement is the ground state of its own series, and lattice dynamics confirm that it is a minimum with respect to atomic displacement as well as to energy. The prediction is testable, and at 3.37 meV/atom above the NaBH4 + Ca(BH4)2 tie-line it sits well inside the metastability range of compounds that have in fact been synthesised.
References
[1] M. A. Hannan, M. M. Hoque, A. Mohamed, A. p. 0 Ayob, Renew. Sustain. Energy Rev. p. 0, doi:10.1016/j/rser-2016-11-17
[2] N. Kittner, F. Lill, D. M. Kammen, Nat. Energy 2, 17
[3] S. Ould Amrouche et al., Int. J. Hydrogen Energy 41, 20914–20927
[4] S. Ismail, G. E. Khedr, F. Z. Salem, N. K. Allam, Energy & Fuels 39, 13703–13714
[5] E. Wolf, in Electrochemical Energy Storage for Renewable Sources and Grid Balancing, Elsevier (2015), pp. p. 1 129–142
[6] M. D. Allendorf et al., Nat. Chem. 14, 1214–1223
[7] H. Liu et al., Green Energy Environ. p. 9, 5, 528–537
[8] A. Z. Arsad et al., Int. J. Hydrogen Energy (2022), doi:10. p. 01016/j/ijhydene-2022-03-208
[9] Y. Xu, Y. Zhou, Y. Li, Z. Ding, Molecules 29, 1767
[10] Y. Nakamori et al., Phys. Rev. B 74, 45126
[11] E. harrak Abdechafik et al., Int. J. Hydrogen Energy 52, 1182–1193
[12] Z. Xiong et al., Nat. Mater. 7, 138–141
[13] J. S. Hummelshøj et al., J. Chem. Phys. 131, 014
[14] V. Ozoliņš, E. H. Majzoub, C Wolverton, J Am Chem Soc., 131, 230–237
[15] Y. Guo et al., J Mol Model., 19, 5135–5142
[16] P. Zhou et al., Adv Mater., 37, e2413430
[17] A. M. Krajewski, J W Siegel, J Xu, Z K Liu, Comput Mater Sci., 208, 11
[18] L Ward et al., Phys. Rev. B 96, 024104
[19] Q Zhou et al., Int J Hydrogen Energy, 179, 15
[20] K T Møller et al., Dalton Trans. 45, 831–840
[21] A Merchant et al., Nature 624, 80–85
[22] S P Ong et al., pymatgen, Comput Mater Sci. p. 4, 68, 314–319
[23] G L W Hart, R W Forcade, Phys. Rev. B 77, 224115
[24] M Ångqvist et al., ICET, Adv Theory Simul. p. 4, 2, 1900015
[25] G L W Hart, L J Nelson, R W Forcade, Comput Mater Sci., 59, 101–107
[26] W L Bragg, E J Williams, Proc. R Soc Lond A 145, 699–730
[27] A R Oganov, M Valle, J Chem Phys. 130, 104504
[28] A van de Walle et al., Calphad 26, 539–553
[29] A van de Walle, G Ceder, Automating First-Principles Phase Diagram Calculations
[30] A van de Walle, Calphad (2009), doi:10.16/j/calphad-2008-12-0
[31] I Batatia, D P Kovács, G N C Simm, C Ortner, G Csányi, arXiv:2206.07697
[32] I Batatia et al., A foundation model for atomistic materials chemistry, arXiv:2401. p. 000096
[33] A Hjorth Larsen et al., J Phys.: Condens. Matter 29, 273002
[34] G Kresse, J Hafner, Phys. Rev. B 47, 558–561
[35] G Kresse, J Furthmüller, Comput Mater Sci., 6, 15–50
[36] G Kresse, J Furthmüller, Phys. Rev. B 54, 11169–11186
[37] P E Blöchl, Phys. Rev. B 50, 17953–17979
[38] G Kresse, D Joubert, Phys. Rev. B 59, 1758–1775
[39] J P Perdew, K Burke, M Ernzerhof, Phys. Rev. Lett., 77, 3865–3868
[40] S Grimme et al., J Chem Phys., 132, 154104
[41] H J Monkhorst, J D Pack, Phys. Rev. B 13, 5188–5192
[42] A Togo et al., Scr Mater., 108, 1–5
[43] A Togo et al., J Phys. Soc. Jpn., 92, 012001
[44] A Togo et al., Scr Mater., 15, 1
[45] A Gadallah, hydrophonokit, Zenodo (2026), doi:10.5281/zenodo.20337121
[46] M Gajdóš et al., Phys. Rev. B 73, 45112
[47] X Gonze, C Lee, Phys. Rev. B 55, 10355
[48] A Togo et al., J Phys.: Condens. Matter 35, 353001
[49] N Verdal et al., J Phys. Chem. C 114, 10027–10033
[50] P Schouwink et al., Nat. Commun., 5, 5706
[51] P Schouwink et al., Energies 8, 8286–8299
[52] C E Patrick et al., Phys. Rev. B 92, 201205(R)
[53] I Pallikara et al., Electron Struct., 4, 33002
[54] M Paskevicius et al., Chem. Soc. Rev., 46, 1565–1634
[55] W Sun et al., Sci. Adv., 2, e1600225
[56] S R H Jensen et al., J Phys. Chem. C 119, 27919–27929
[57] M J Buerger, J Chem Phys., 15, 1–16 (1947)
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Data-driven candidate search refinement: The AI system can perform an unbiased structural search by generating 2238 candidates across both perovskite and non-perovskite topologies, ensuring that
the candidate set would not be limited to the topology the analogues happen to share.
-
Structure selection hierarchy: The system can prioritize structures based on a multi-stage validation process, moving from an initial machine learning potential screening to full first principles relaxation, as
the most stable structure is not a perovskite
andG2 is deeper than all of them at -0.3469 eV/atom.
-
Cation ordering resolution: The AI can resolve the cation arrangement degree of freedom by performing exhaustive enumeration across different supercell variations and swap fractions, as this process
controls all three [sources] [of error].
-
Potential application division: The system can learn to distinguish between the roles of machine learning potentials and first-principles calculations, recognizing that
a machinelearning potential is therefore the right instrument for covering a candidate library exhaustively and the wrong one for deciding between its leading members.
-
Predictive testing capability: The AI can generate simulated diffraction patterns from relaxed cells to test the prediction, as this
is offered as the means of testing that prediction, not as evidence that it holds,
allowing researchers to check ifthe predicted framework is identified by seven reflections that neither parent phase nor the competing framework produces.
-
Stability assessment: The system can assess dynamical stability by analyzing phonon spectra to confirm a structure's viability, noting that for G2,
its ordered arrangement is the ground state of its own series
and it hasno imaginary mode at any wavevector the supercell resolves exactly.
Sources
- MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields
- A foundation model for atomistic materials chemistry
- $\texttt{Spglib}$: a software library for crystal symmetry search
Related papers
- Structural contribution to light-induced gap suppression in Ta 2 NiSe 5
- Anatomy of fast current-induced skyrmion motion in synthetic antiferromagnets
- Magnetic control of an exciton-polariton condensate in a van der Waals magnet
- High-pressure elastic properties of GeO2 polymorphs up to 120 GPa
- Decoupling Strain-Rate Sensitivity and Deformation Length Scale Effects in Neutron-Irradiated Tungsten: A Coupled Nano-Indentation, HR-EBSD and Crystal Plasticity Study
- Magnetism modulation in Co3Sn2S2 by current-assisted domain wall motion