Pith. sign in

REVIEW 3 major objections 2 minor 1 cited by

Full-cycle device-scale simulations of memory materials with a tailored atomic-cluster-expansion potential

T0 review · 3 major / 2 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that an optimized atomic-cluster-expansion (ACE) machine-learned potential, GST-ACE-24, makes full-cycle device-scale simulations of Ge-Sb-Te phase-change memory cells practical on CPU clusters, reproducing RESET, SET…

desk verdict A real first: full RESET-SET-RESET device-scale simulations with an ACE potential, solid and useful, with validation caveats that are addressable. read the letter →

arxiv 2502.08393 v1 pith:T3SWDKN6 submitted 2025-02-12 cond-mat.mtrl-sci physics.comp-ph

Yuxing Zhou , Daniel F. Thomas du Toit , Stephen R. Elliott , Wei Zhang , Volker L. Deringer This is my paper · ORCID
classification cond-mat.mtrl-sciphysics.comp-ph
keywords phase-changematerialsatomicclusterexpansionmachine-learnedinteratomicpotentialsGe-Sb-Temoleculardynamicsdevice-scalesimulationcrystallizationkineticsneuromorphiccomputing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper aims to remove the last computational bottleneck in atomistic simulations of phase-change memory (PCM) devices: the slow crystallization step that writes a digital 'one'. Earlier machine-learned potentials, built with the Gaussian-approximation-potential (GAP) framework, could simulate the fast RESET (melting-and-quench) step on device-scale models but would have consumed more than 150 million CPU core hours for a single 10-ns SET run. The authors construct a new potential, GST-ACE-24, using the atomic cluster expansion (ACE), re-fitting an expanded DFT dataset with optimized hyperparameters and adding small hard-sphere-random and RSS-derived structures for robustness. They report that the resulting potential is more than 400 times more efficient than the previous GAP model on CPU-only high-performance computers, memory-efficient enough for billion-atom simulations, and capable of simulating complete RESET-SET-RESET cycles for cross-point cells and multi-state programming for mushroom-type cells. If the potential's nucleation kinetics are quantitatively right, this makes device-scale, multiple-cycle simulation a practical tool for studying memory and neuromorphic devices.

What carries the argument

The atomic cluster expansion (ACE) is the central engine: the local environment of every atom is expanded in a many-body basis, expressed through radial functions and spherical harmonics, coupled into invariant 'B-basis' functions, and the per-atom energy is evaluated by simple summations, avoiding the Gaussian-process regression that makes GAP-based models slow. This switch accounts for the more-than-400-fold CPU speed-up and the much smaller memory footprint (up to about 650 million atoms on half a terabyte of RAM), which is what brings tens-of-nanosecond crystallization within reach. Equally load-bearing is the data pipeline that makes the model robust: along with melt-quench and phase-transition configurations, the training set includes 906 small random structures (hard-sphere random and ACE-driven random-structure-search structures), which the ablation studies show are necessary to prevent unphysical short contacts and lost atoms during long MD runs; the hyperparameters were optimized with a dedicated cross-platform tool cited in the paper.

What would settle it

Run many independent polymorphic crystallization simulations of amorphous Ge1Sb2Te4 at 600 K with GST-ACE-24 and compare the resulting nucleation rates, incubation times, and grain-size distributions with AIMD-based benchmarks (e.g., metadynamics or seeded-crystal estimates) or with experimental time-temperature-transformation data; a discrepancy of more than an order of magnitude in the nucleation rate would show that the potential does not provide a quantitatively accurate SET picture. A cheaper falsifier is to repeat the reported single 20-ns SET trajectory several times and verify that the grain count and the crystal-like atom fraction are reproducible rather than a single stochastic realization.

Watch

Extended reading notes

Core claim

The central claim is that a deliberately engineered ACE potential, GST-ACE-24, enables full-cycle, device-scale molecular dynamics of Ge-Sb-Te phase-change materials on CPU-only HPC hardware, without any enhanced sampling or pre-seeded nuclei. Using a 532,980-atom cross-point cell, the authors reproduce a 10-ps non-isothermal RESET, then anneal the amorphous cell at 600 K for 20 ns and observe spontaneous homogeneous nucleation, yielding 277 randomly oriented grains with an average diameter of about 4.6 nm; a second, weaker 40-ps RESET pulse then melts this polycrystalline state, which is easier to melt than the stable trigonal phase. For a 794,808-atom mushroom-type cross-section, they create two amorphous zones of different diameters and show that subsequent crystallization at 600 K involves a competition between interface growth and bulk nucleation, with the nucleation share rising from 46% to 65% as the amorphous region grows. The authors also validate structural properties (radial and angular distributions, homopolar bonds, tetrahedral fraction, bond-length asymmetry) across seven compositions along the GeTe-Sb2Te3 tie-line, indicating chemical transferability and the ability to describe ageing-related structural features.

Load-bearing premise

The load-bearing premise is that GST-ACE-24 quantitatively captures the homogeneous nucleation and growth kinetics of amorphous Ge1Sb2Te4 at 600 K and comparable temperatures; if its free-energy barriers are slightly off, the observed 277-grain polycrystalline SET state and the growth-versus-nucleation balance could change qualitatively.

Editorial extensions

If this is right

  • Full RESET-SET-RESET cycling of cross-point PCM cells becomes a standard, repeatable simulation on CPU clusters, enabling studies of structural and compositional evolution over many write-erase cycles.
  • Unseeded SET simulations become possible, turning homogeneous nucleation, grain statistics, and polycrystalline melting into directly observable quantities rather than inferred ones.
  • In mushroom-type cells, multi-level (analogue) programming can be simulated atomistically, which is relevant for neuromorphic computing; the observed growth-nucleation competition links RESET-pulse amplitude to the recrystallization pathway.
  • Because the model also captures local bond-length asymmetry, the same potential can address resistance-drift-related aging and cycle-to-cycle stochasticity in one device model.
  • The efficiency gain and memory reduction imply that device-scale simulations are no longer limited to the fastest process (RESET) but can cover the full programming cycle.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the quantitative validity of the SET picture rests on nucleation kinetics that the paper does not benchmark; a natural extension is to compute nucleation rates with the potential and compare against experimental time-temperature-transformation data or AIMD-based free-energy estimates.
  • Editorial inference: because the SET trajectory is a single run, the reported 277-grain outcome is one stochastic realization; repeated independent SET runs would quantify cycle-to-cycle variability and connect directly to the measured device-to-device variation in PCM cells.
  • Editorial inference: the same ACE-plus-random-structures recipe could be ported to other chalcogenide systems, such as Sb-Te alloys or Ovonic threshold switching (OTS) selector materials, to build device-scale simulations of the full selector-plus-memory stack.
  • Editorial inference: the energy estimate (2,500 kWh for the full cycle) highlights that routine multi-cycle simulations are affordable on current supercomputers, but large-scale adoption might be accelerated by combining ACE with enhanced sampling or coarse-graining for the slow nucleation step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 2 minor

Summary. This manuscript reports a new machine-learned interatomic potential, GST-ACE-24, for Ge-Sb-Te phase-change materials, built from the GST-GAP-22 dataset with XPOT hyperparameter optimization and iterative dataset expansion including hard-sphere random structures and ACE-driven random structure searching. The authors demonstrate large speedups over GAP on CPU architectures, memory efficiency up to billion-atom systems, high-temperature annealing robustness, and systematic ablation studies on random structures, atomic properties, and basis functions. They then present full-cycle RESET-SET-RESET simulations for a 532,980-atom cross-point model and partial programming for a 794,808-atom mushroom-type model, including spontaneous nucleation at 600 K over 20 ns and a growth-versus-nucleation competition analysis. The central claim is that this potential makes device-scale, full-cycle simulations of PCM memory cells feasible on CPU-only high-performance computing systems.

Significance. If the quantitative kinetic content of the simulations is accepted, this is a substantial advance: complete programming cycles of realistic PCM device geometries would become accessible to atomistic simulation, and the reported efficiency benchmarks (more than 400x speedup over GAP, memory limits, scaling behavior) are valuable and largely convincing. The paper is unusually thorough in reporting robustness tests and ablation studies, and the structural validation across seven GST compositions along the GeTe-Sb2Te3 tie-line is a clear strength. However, the central device-scale demonstration rests on the potential's ability to describe homogeneous nucleation kinetics, and that ability is not directly validated against AIMD, enhanced-sampling, or experimental rate data. In addition, the reported RMSE values are computed on a test set that shares its origins with the training data. These issues need to be addressed before the strong claims can be fully endorsed.

major comments (3)
  1. [Results: Full-cycle operations for cross-point GST memory devices (Fig. 3c)] The SET leg of the full cycle is supported by a single 20 ns NVT trajectory in which 277 grains nucleate in amorphous Ge1Sb2Te4. Homogeneous nucleation is exponentially sensitive to the amorphous/crystalline free-energy difference and the interfacial free energy, and a single stochastic trajectory cannot distinguish a potential with correct nucleation kinetics from one that nucleates too easily. The paper does not benchmark nucleation rate, incubation time, or crystal-growth velocity against AIMD, enhanced-sampling results, or experiment. I request additional evidence: at least several independent SET runs from different random seeds to establish run-to-run variability, and quantitative comparisons of nucleation rate and growth velocity with literature values or with AIMD/meta-dynamics on smaller cells. Without such benchmarks, the central claim of a physically meaningful full-cycle SET simulation is not established.
  2. [Methods: The GST-ACE-24 potential; Fig. 2c] The testing dataset used to guide the XPOT optimization and to report RMSE values in Fig. 2c consists of conventional disordered structures from Ref. 31 and intermediate phase-transition configurations from Ref. 18, as stated in the Methods. These are the same sources that supply training structures: Supplementary Table 1 lists melt-quenched and phase-transition configurations from Ref. 31, and iter-1 to iter-3 add structures of the same classes. The reported RMSEs are therefore at least partially in-sample and are likely over-optimistic as measures of transferability to spontaneous crystallization. Please provide an out-of-sample test set, for example structures from compositions or thermodynamic conditions not represented in training or a source-based split, and report the degree of overlap between the training and test sets.
  3. [Results: Full-cycle operations for in-memory computing (Fig. 4e-g)] The competition between growth-driven and nucleation-driven recrystallization (54%/46% for state I and 35%/65% for state II) is derived from single trajectories. Given the stochasticity of nucleation, these percentages and the conclusion that SET speed is almost independent of the size of the amorphized region need statistical support from multiple independent runs. The authors should either provide error bars from repeated simulations or temper the quantitative conclusion accordingly.
minor comments (2)
  1. [Methods: Molecular-dynamics simulations] The sentence 'Additional energy was added to the kinetic energy of the atoms in the programming regions, with a timestep of 2 ps' appears to contain a typo; the following sentence and all other statements in the paper indicate a timestep of 2 fs.
  2. [Fig. 2c caption] The caption states that RMSE values were calculated on the testing dataset but does not mention that this testing dataset is drawn from the same sources as the training data; this should be stated explicitly in the caption or in the main text for transparency.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor validation circularity from XPOT test-set reuse; the central device-scale simulations are genuine MD runs and are not reduced to fits.

  1. fitted input called prediction [Methods, 'The GST-ACE-24 potential' (XPOT optimisation paragraph); Supplementary Note 2.]
    "To guide the target of the XPOT optimisation, we defined a testing dataset consisting of conventional disordered structures (≈ 200 atoms each) and intermediate configurations during phase transitions (1,008 atoms each). These two types of structures were taken from AIMD simulations reported in Ref. 31 and Ref. 18, respectively. This testing dataset was also used in the computation of RMSE values shown in Fig. 2c."

    The same testing dataset is used both as the objective for hyperparameter optimisation and as the benchmark on which the reported RMSE values in Fig. 2c are computed. The training data for GST-ACE-24 are built from the same sources: the GST-GAP-22 dataset from Ref. 31 and additional disordered AIMD configurations from Ref. 31, as detailed in Supplementary Table 1. Therefore the testing RMSE is not an independent estimate; the hyperparameters were selected to minimise error on configurations of the same origin as the training set. This is a selection-on-the-test-set circularity. It does not, however, make the full-cycle device simulations themselves circular, because the SET/RESET observables are generated by MD trajectories with the fitted potential and were not used as fitting targets.

full rationale

The central derivation chain in this paper is a standard machine-learned-potential workflow: DFT-labelled configurations are used to fit GST-ACE-24, and the fitted potential is then used in NVE/NVT molecular dynamics to simulate RESET, SET, and a second RESET. None of the reported device-scale outcomes, such as the 277 crystalline grains, the 4.6 nm average grain diameter, the 54/46 growth-versus-nucleation split, or the melting behaviour, was used as a fitting target; these are genuine outputs of the simulations. The computational-efficiency claims are direct measurements on ARCHER2 and GPU hardware, not fitted quantities. The one identifiable circular element is the validation procedure: the testing dataset from Refs. 31 and 18 is used both to guide XPOT hyperparameter optimisation and to compute the reported RMSE values, while the same references also provide training structures, so the Fig. 2c accuracy numbers are not fully independent. This is a moderate methodological caveat, but it is not load-bearing for the central full-cycle demonstration, which would stand or fall on the physical transferability of the potential rather than on the specific RMSE values. The structural validations against AIMD are partly in-sample but still show that the ACE potential remains stable and structurally reasonable in MD. The lack of a dedicated nucleation-rate benchmark is a correctness risk, not a circularity, so it does not raise the score. Overall, the paper is essentially self-contained against external DFT data, and the observed test-set overlap is minor.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claims rest on the accuracy and robustness of a fitted ML potential. Free parameters include ACE hyperparameters, model complexity choices, and dataset composition. No new physical entities, particles, or forces are introduced. The key assumptions are the reliability of PBE-DFT reference data, the completeness of the ACE basis, and the idealized thermal and structural models used for device simulations, including the representativeness of single MD trajectories.

free parameters (4)
  • XPOT-optimized ACE hyperparameters (cut-off, radial parameter, k2, k3) = 8.0 Å, 10, 4.577, 0.101
    Optimized using Bayesian optimization to minimize RMSE on a testing dataset; the testing dataset may overlap with the training data, so these are effectively fitted to the target data.
  • Model complexity (number of basis functions, number of atomic properties P) = 3000 basis functions, P=3
    Chosen by the authors based on ablation studies trading accuracy and speed; these are hand-picked design parameters.
  • Training dataset composition: number of random structures and iterative additions = 906 random structures (319 hard-sphere + 587 ACE-RSS); 5 iterations
    Determined by the need for MD robustness; the specific counts were the result of iterative testing, not a formal fit.
  • Relative energy weighting in final upfit = not quoted
    Increased relative weighting of the energy was applied after XPOT; the exact value is a tuneable choice.
assumptions (6)
  • domain assumption PBE-DFT with the given pseudopotentials provides accurate reference energies and forces for GST alloys.
    Used to label the training and testing data; if PBE is inaccurate for the relevant states, the potential inherits the error.
  • standard math The atomic cluster expansion (ACE) framework is sufficiently complete and stable to represent the GST potential energy surface.
    Relies on published completeness and stability results (Refs. 39-41).
  • domain assumption A fixed amorphous GST slab acts as a realistic thermal barrier at the periodic boundary.
    Used in both device models; real devices have different electrode/barrier materials (e.g., TiN, OTS), so the thermal boundary conditions are idealized.
  • domain assumption Non-isothermal heating by injecting kinetic energy into atoms, then removing it gradually, adequately represents RESET pulse heating.
    This protocol was introduced in Ref. 31 and used here; it does not include Joule heating, electrical transport, or thermal diffusion, and the authors note it gives a theoretical lower bound for energy.
  • domain assumption The SOAP-based crystallinity measure k-bar with a cutoff of 0.57 is a valid order parameter for classifying crystal-like vs amorphous-like atoms.
    Adopted from Ref. 18; the threshold is transferred from that work without recalibration here.
  • domain assumption A single MD trajectory is representative of the nucleation and crystallization ensemble at 600 K.
    The reported grain counts and growth/nucleation fractions come from single runs; no repeated trajectories or statistical error bars are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Full-cycle device-scale simulations of memory materials with a tailored atomic-cluster-expansion potential." pith.science (2026). https://pith.science/paper/T3SWDKN6

@misc{pith2026250208393,
  author       = {Pith},
  title        = {Pith review of: Full-cycle device-scale simulations of memory materials with a tailored atomic-cluster-expansion potential},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T3SWDKN6}},
  note         = {Machine review of arXiv:2502.08393}
}
read the original abstract

Computer simulations have long been key to understanding and designing phase-change materials (PCMs) for memory technologies. Machine learning is now increasingly being used to accelerate the modelling of PCMs, and yet it remains challenging to simultaneously reach the length and time scales required to simulate the operation of real-world PCM devices. Here, we show how ultra-fast machine-learned interatomic potentials, based on the atomic cluster expansion (ACE) framework, enable simulations of PCMs reflecting applications in devices with excellent scalability on high-performance computing platforms. We report full-cycle simulations -- including the time-consuming crystallisation process (from digital "zeroes" to "ones") -- thus representing the entire programming cycle for cross-point memory devices. We also showcase a simulation of full-cycle operations, relevant to neuromorphic computing, in a mushroom-type device geometry. Our work provides a springboard for the atomistic modelling of PCM-based memory and neuromorphic computing devices -- and, more widely, it illustrates the power of highly efficient ACE ML models for materials science and engineering.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Machine-learning-driven modelling of amorphous and polycrystalline BaZrS$_{3}$

    cond-mat.mtrl-sci 2025-06 conditional novelty 6.0 of 10

    A machine-learned interatomic potential for BaZrS3 models amorphous, grain-boundary, and polycrystalline structures, with simulated XRD, PDF, and EXAFS matching experiment.

Reference graph

Works this paper leans on

98 extracted references · 71 canonical work pages · cited by 1 Pith paper

  1. [1]

    Zhang, W., Mazzarello, R., Wuttig, M. & Ma, E. Designing crystallization in phase-change materials for universal memory and neuro -inspired computing. Nat. Rev. Mater. 4, 150– 168 (2019)

    2019

  2. [2]

    D., Bhaskaran, H

    Feldmann, J., Youngblood, N., Wright, C. D., Bhaskaran, H. & Pernice, W. H. P. All-opti- cal spiking neurosynaptic networks with self -learning capabilities. Nature 569, 208–214 (2019)

    2019

  3. [3]

    & Meyer, R

    Pellizzer, F., Pirovano, A., Bez, R. & Meyer, R. L. Status and perspectives of chalcogenide- based CrossPoint memories. In 2023 International Electron Devices Meeting (IEDM) 1–4 (2023). DOI: 10.1109/IEDM45741.2023.10413669

    arXiv 2023

  4. [4]

    Ambrogio, S. et al. An analog-AI chip for energy -efficient speech recognition and tran- scription. Nature 620, 768–775 (2023)

    2023

  5. [5]

    Le Gallo, M. et al. A 64-core mixed-signal in-memory compute chip based on phase-change memory for deep neural network inference. Nat. Electron. 6, 680–693 (2023)

    2023

  6. [6]

    & Zhang, W

    Zhou, W., Shen, X., Yang, X., Wang, J. & Zhang, W. Fabrication and integration of pho- tonic devices for phase -change memory and neuromorphic computing. Int. J. Extrem. Manuf. 6, 022001 (2024)

    2024

  7. [7]

    & Yamada, N

    Wuttig, M. & Yamada, N. Phase-change materials for rewriteable data storage. Nat. Mater. 6, 824–832 (2007)

    2007

  8. [8]

    & Eleftheriou, E

    Tuma, T., Pantazi, A., Le Gallo, M., Sebastian, A. & Eleftheriou, E. Stochastic phase - change neurons. Nat. Nanotechnol. 11, 693–699 (2016)

    2016

Show all 98 references
  1. [9]

    & Jones, R

    Akola, J. & Jones, R. Structural phase transitions on the nanoscale: The crucial pattern in the phase-change materials Ge2Sb2Te5 and GeTe. Phys. Rev. B 76, 235201 (2007)

    2007

  2. [10]

    D., Krack, M

    Caravati, S., Bernasconi, M., Kühne, T. D., Krack, M. & Parrinello, M. Coexistence of tetrahedral- and octahedral -like sites in amorphous phase change materials. Appl. Phys. Lett. 91, 171906 (2007)

    2007

  3. [11]

    Xu, M., Cheng, Y., Sheng, H. & Ma, E. Nature of atomi c bonding and atomic structure in the phase-change Ge2Sb2Te5 Glass. Phys. Rev. Lett. 103, 195502 (2009)

    2009

  4. [12]

    & Robertson, J

    Huang, B. & Robertson, J. Bonding origin of optical contrast in phase-change memory ma- terials. Phys. Rev. B 81, 081204(R) (2010)

    2010

  5. [13]

    Raty, J.-Y. et al. A quantum‐mechanical map for bonding and properties in solids. Adv. Mater. 31, 1806280 (2019)

    2019

  6. [14]

    Wang, X.-D. et al. Multiscale simulations of growth-dominated Sb2Te phase-change mate- rial for non-volatile photonic applications. npj Comput. Mater. 9, 136 (2023)

    2023

  7. [15]

    & Zhang, W

    Shen, X., Chu, R., Jiang, Y. & Zhang, W. Progress on materials design and multiscale sim- ulations for phase-change memory. Acta Metall. Sin. 60, 1362–1378 (2024). 25

    2024

  8. [16]

    & Elliott, S

    Hegedüs, J. & Elliott, S. R. Microscopic origin of the fast crystallization ability of Ge -Sb- Te phase-change memory materials. Nat. Mater. 7, 399–405 (2008)

    2008

  9. [17]

    & Jones, R

    Kalikka, J., Akola, J. & Jones, R. O. Crystallization processes in the phase change material Ge2Sb2Te5: Unbiased density functional/molecular dynamics simulations. Phys. Rev. B 94, 134105 (2016)

    2016

  10. [18]

    Xu, Y. et al. Unraveling crystallization mechanisms and electronic structure of phase- change materials by large-scale ab initio simulations. Adv. Mater. 34, 2109139 (2022)

    2022

  11. [19]

    First principles neural network potentials for reactive simulations of large molec- ular and condensed systems

    Behler, J. First principles neural network potentials for reactive simulations of large molec- ular and condensed systems. Angew. Chem. Int. Ed. 56, 12828–12840 (2017)

    2017

  12. [20]

    L., Caro, M

    Deringer, V. L., Caro, M. A. & Csányi, G. Machine learning interatomic potentials as emerging tools for materials science. Adv. Mater. 31, 1902765 (2019)

    2019

  13. [21]

    & Aspuru -Guzik, A

    Friederich, P., Häse, F., Proppe, J. & Aspuru -Guzik, A. Machine -learned potentials for next-generation matter simulations. Nat. Mater. 20, 750–761 (2021)

    2021

  14. [22]

    C., Miceli, G., Caravati, S., Behler, J

    Sosso, G. C., Miceli, G., Caravati, S., Behler, J. & Bernasconi, M. N eural network intera- tomic potential for the phase change material GeTe. Phys. Rev. B 85, 174103 (2012)

    2012

  15. [23]

    & Parrinello, M

    Behler, J. & Parrinello, M. Generalized neural-network representation of high-dimensional potential-energy surfaces. Phys. Rev. Lett. 98, 146401 (2007)

    2007

  16. [24]

    C., Salvalaglio, M., Behler, J., Bernasconi, M

    Sosso, G. C., Salvalaglio, M., Behler, J., Bernasconi, M. & Parrinello, M. Heterogeneous crystallization of the phase change material GeTe via atomistic simulations. J. Phys. Chem. C 119, 6428–6434 (2015)

    2015

  17. [25]

    & Bernasconi, M

    Abou El Kheir, O., Bonati, L., Parrinello, M. & Bernasconi, M. Unraveling the crystalliza- tion kinetics of the Ge 2Sb2Te5 phase change compound with a machine-learned interatomic potential. npj Comput. Mater. 10, 33 (2024)

    2024

  18. [26]

    Gabardi, S. et al. Atomistic simulations of the crystallization and aging of GeTe nanowires. J. Phys. Chem. C 121, 23827–23838 (2017)

    2017

  19. [27]

    Mocanu, F. C. et al. Modeling the phase-change memory material, Ge2Sb2Te5, with a ma- chine-learned interatomic potential. J. Phys. Chem. B 122, 8998–9006 (2018)

    2018

  20. [28]

    & Bernasconi, M

    Dragoni, D., Behler, J. & Bernasconi, M. Mechanism of amorphous phase stabilization in ultrathin films of monoatomic phase change material. Nanoscale 13, 16146–16155 (2021)

    2021

  21. [29]

    Mo, P. et al. Accurate and efficient molecular dynamics based on machine learning and non von Neumann architecture. npj Comput. Mater. 8, 107 (2022)

    2022

  22. [30]

    P., Payne, M

    Bartók, A. P., Payne, M. C., Kondor, R. & Csányi, G. Gaussian approximation potentials: the accuracy of quantum mechanics, without the electrons. Phys . Rev. Lett. 104, 136403 (2010)

    2010

  23. [31]

    & Deringer, V

    Zhou, Y., Zhang, W., Ma, E. & Deringer, V. L. Device-scale atomistic modelling of phase- change memory materials. Nat. Electron. 6, 746–754 (2023). 26

    2023

  24. [32]

    & Sun, Z

    Li, K., Liu, B., Zhou, J. & Sun, Z. Revealing the crystallization dynamics of Sb –Te phase change materials by large-scale simulations. J. Mater. Chem. C 12, 3897–3906 (2024)

    2024

  25. [33]

    Batzner, S. et al. E(3)-equivariant graph neural networks for data -efficient and accurate interatomic potentials. Nat. Commun. 13, 2453 (2022)

    2022

  26. [34]

    P., Simm, G

    Batatia, I., Kovács, D. P., Simm, G. N. C., Ortner, C. & Csányi, G. MACE: higher order equivariant message passing neural networks for fast and accurate force fields. Preprint at arXiv:2206.07697 (2023)

    2023 arXiv

  27. [35]

    & Sun, Z

    Wang, G., Sun, Y., Zhou, J. & Sun, Z. PotentialMind: graph convolutional machine learn- ing potential for Sb–Te binary compounds of multiple stoichiometries. J. Phys. Chem. C 127, 24724–24733 (2023)

    2023

  28. [36]

    L., Katti, K

    Chang, C., Deringer, V. L., Katti, K. S., Van Speybroeck, V. & Wolverton, C. M. Simula- tions in the era of exascale computing. Nat. Rev. Mater. 8, 309–313 (2023)

    2023

  29. [37]

    Atomic cluster expansion for accurate and transferable interatomic potentials

    Drautz, R. Atomic cluster expansion for accurate and transferable interatomic potentials. Phys. Rev. B 99, 014104 (2019)

    2019

  30. [38]

    R., Arbaugh, T

    Dunton, O. R., Arbaugh, T. & Starr, F. W. Computationally effic ient machine -learned model for GST phase change materials via direct and indirect learning. J. Chem. Phys. 162, 034501 (2025)

    2025

  31. [39]

    Lysogorskiy, Y. et al. Performant implementation of the atomic cluster expansion (PACE) and application to copper and silicon. npj Comput. Mater. 7, 97 (2021)

    2021

  32. [40]

    Bochkarev, A. et al. Efficient parametrization of the atomic cluster expansion. Phys. Rev. Mater. 6, 013804 (2022)

    2022

  33. [41]

    Dusson, G. et al. Atomic cluster expansion: Completeness, efficiency and stability. J. Com- put. Phys. 454, 110946 (2022)

    2022

  34. [42]

    Deringer, V. L. et al. Gaussian process regression for materials and molecules. Chem. Rev. 121, 10073–10141 (2021)

    2021

  35. [43]

    & Drautz, R

    Qamar, M., Mrovec, M., Lysogorskiy, Y., Bochkarev, A. & Drautz, R. Atomic cluster ex- pansion for quantum-accurate large-scale simulations of carbon. J. Chem. Theory Comput. 19, 5151–5167 (2023)

    2023

  36. [44]

    R., Toit, D

    Zhou, Y., Elliott, S. R., Toit, D. F. T. du, Zhang, W. & Deringer, V. L. The pathway to chirality in elemental tellurium. Preprint at arXiv:2409.03860 (2024)

    2024 arXiv

  37. [45]

    C., Rohrer, J., Albe, K

    Erhard, L. C., Rohrer, J., Albe, K. & Deringer, V. L. Modelling atomic and nanoscale struc- ture in the silicon–oxygen system through active machine learning. Nat. Commun. 15, 1927 (2024)

    2024

  38. [46]

    Morrow, J. D. & Deringer, V. L. Indirect learning and physically guided validation of in- teratomic potential models. J. Chem. Phys. 157, 104105 (2022)

    2022

  39. [47]

    P., Burke, K

    Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 77, 3865–3868 (1996). 27

    1996

  40. [48]

    Thomas du Toit, D. F. & Deringer, V. L. Cross -platform hyperparameter optimization for machine learning interatomic potentials. J. Chem. Phys. 159, 024803 (2023)

    2023

  41. [49]

    F., Zhou, Y

    Thomas du Toit, D. F., Zhou, Y. & Deringer, V. L. Hyperparameter optimization for atomic cluster expansion potentials. J. Chem. Theory Comput. 20, 10103–10113 (2024)

    2024

  42. [50]

    Pickard, C. J. & Needs, R. J. High- pressure phases of silane. Phys. Rev. Lett. 97, 045504 (2006)

    2006

  43. [51]

    Pickard, C. J. & Needs, R. J. Ab initio random structure searching. J. Phys.: Condens. Mat- ter 23, 053201 (2011)

    2011

  44. [52]

    L., Proserpio, D

    Deringer, V. L., Proserpio, D. M., Csányi, G. & Pickard, C. J. Data -driven learning and prediction of inorganic crystal structures. Faraday Discuss. 211, 45–59 (2018)

    2018

  45. [53]

    L., Pickard, C

    Deringer, V. L., Pickard, C. J. & Csányi, G. Data-driven learning of total and local energies in elemental boron. Phys. Rev. Lett. 120, 156001 (2018)

    2018

  46. [54]

    & Deringer, V

    Bernstein, N., Csányi, G. & Deringer, V. L. De novo exploration and self -guided learning of potential-energy surfaces. npj Comput. Mater. 5, 99 (2019)

    2019

  47. [55]

    Batatia, I. et al. The design space of E(3)-equivariant atom-centred interatomic potentials. Nat. Mach. Intell. 7, 56–67 (2025)

    2025

  48. [56]

    & Margraf, J

    Stocker, S., Gasteiger, J., Becker, F., Günnemann, S. & Margraf, J. T. How robust are mod- ern graph neural network potentials in long and hot molecular dynamics simulations? Mach. Learn.: Sci. Technol. 3, 045010 (2022)

    2022

  49. [57]

    Unke, O. T. et al. Machine learning force fields. Chem. Rev. 121, 10142–10186 (2021)

    2021

  50. [58]

    Ben Mahmoud, C., Gardner, J. L. A. & Deringer, V. L. Data as the next challenge in atom- istic machine learning. Nat. Comput. Sci. 4, 384–387 (2024)

    2024

  51. [59]

    Allen, A. E. A. et al. Learning together: Towards foundation models for machine learning interatomic potentials with meta-learning. npj Comput. Mater. 10, 154 (2024)

    2024

  52. [60]

    Liu, Y. et al. An automated framework for exploring and learning potential-energy surfaces. Preprint at arXiv:2412.16736 (2024)

    2024 arXiv

  53. [61]

    L., Caro, M

    Deringer, V. L., Caro, M. A. & Csányi, G. A general-purpose machine-learning force field for bulk and nanostructured phosphorus. Nat. Commun. 11, 5461 (2020)

    2020

  54. [62]

    Pickard, C. J. Ephemeral data derived potentials for random structure search. Phys. Rev. B 106, 014102 (2022)

    2022

  55. [63]

    Pickard, C. J. Beyond theory-driven discovery: introducing hot random search and datum- derived structures. Faraday Discuss. 256, 61–84 (2025)

    2025

  56. [64]

    P., Kondor, R

    Bartók, A. P., Kondor, R. & Csányi, G. On representing chemical environments. Phys. Rev. B 87, 184115 (2013). 28

    2013

  57. [65]

    & Wuttig, M

    Wełnic, W. & Wuttig, M. Reversible switching in phase -change materials. Mater. Today 11, 20–27 (2008)

    2008

  58. [66]

    Loke, D. et al. Breaking the speed limits of phase -change memory. Science 336, 1566– 1569 (2012)

    2012

  59. [67]

    Loke, D. K. et al. Ultrafast nanoscale phase-change memory enabled by single-pulse con- ditioning. ACS Appl. Mater. Interfaces 10, 41855–41860 (2018)

    2018

  60. [68]

    & Sun, Z

    Li, Z., Si, C., Zhou, J., Xu, H. & Sun, Z. Yttrium -doped Sb2Te3: a promising material for phase-change memory. ACS Appl. Mater. Interfaces 8, 26126–26134 (2016)

    2016

  61. [69]

    Rao, F. et al. Reducing the stochasticity of crystal nucleation to enable subnanosecond memory writing. Science 358, 1423–1427 (2017)

    2017

  62. [70]

    Wang, Y. et al. Scandium doped Ge 2Sb2Te5 for high-speed and low -power-consumption phase change memory. Appl. Phys. Lett. 112, 133104 (2018)

    2018

  63. [71]

    Hu, S., Xiao, J., Zhou, J., Elliott, S. R. & Sun, Z. Synergy effect of co- doping Sc and Y in Sb2Te3 for phase-change memory. J. Mater. Chem. C 8, 6672–6679 (2020)

    2020

  64. [72]

    Wang, X.-P. et al. Time-dependent density-functional theory molecular-dynamics study on amorphization of Sc-Sb-Te alloy under optical excitation. npj Comput. Mater. 6, 31 (2020)

    2020

  65. [73]

    & Mazzarello, R

    Ronneberger, I., Zhang, W., Eshet, H. & Mazzarello, R. Crystallization properties of the Ge2Sb2Te5 phase-change compound from advanced simulations. Adv. Funct. Mater. 25, 6407–6413 (2015)

    2015

  66. [74]

    & Parrinello, M

    Laio, A. & Parrinello, M. Escaping free-energy minima. Proc. Natl. Acad. Sci. U. S. A. 99, 12562–12566 (2002)

    2002

  67. [75]

    & Jones, R

    Kalikka, J., Akola, J., Larrucea, J. & Jones, R. O. Nucleus -driven crystallization of amor- phous Ge2Sb2Te5: A density functional study. Phys. Rev. B 86, 144113 (2012)

    2012

  68. [76]

    Zhang, B. et al. Vacancy structures and melting behavior in rock-salt GeSbTe. Sci. Rep. 6, 25453 (2016)

    2016

  69. [77]

    https://docs.archer2.ac.uk/user-guide/energy/

    Energy use and emissions. https://docs.archer2.ac.uk/user-guide/energy/

  70. [78]

    Song, Z. et al. 12-state multi-level cell storage implemented in a 128 Mb phase change memory chip. Nanoscale 13, 10455–10461 (2021)

    2021

  71. [79]

    & Bernasconi, M

    Abou El Kheir, O. & Bernasconi, M. Million- atom simulation of the set process in phase change memories at the real device scale. Preprint at arXiv:2501.07384 (2025)

    2025 arXiv

  72. [80]

    M., Schmidt, S

    Larsen, P. M., Schmidt, S. & Schiøtz, J. Robust structural identification via polyhedral tem- plate matching. Model. Simul. Mater. Sci. Eng. 24, 055007 (2016)

    2016

  73. [81]

    Visualization and analysis of atomistic simulation data with OVITO –the Open Visualization Tool

    Stukowski, A. Visualization and analysis of atomistic simulation data with OVITO –the Open Visualization Tool. Model. Simul. Mater. Sci. Eng. 18, 015012 (2010). 29

    2010

  74. [82]

    Nandakumar, S. R. et al. A phase-change memory model for neuromorphic computing. J. Appl. Phys. 124, 152135 (2018)

    2018

  75. [83]

    Sebastian, A. et al. Tutorial: Brain-inspired computing using phase -change memory de- vices. J. Appl. Phys. 124, 111101 (2018)

    2018

  76. [84]

    Zhang, W. & Ma, E. Unveiling the structural origin to control resistance drift in phase - change memory materials. Mater. Today 41, 156–176 (2020)

    2020

  77. [85]

    Raty, J. Y. et al. Aging mechanisms in amorphous phase-change materials. Nat. Commun. 6, 7467 (2015)

    2015

  78. [86]

    Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023)

    2023

  79. [87]

    Cheng, B., Mazzola, G., Pickard, C. J. & Ceriotti, M. Evidence for supercritical behaviour of high-pressure liquid hydrogen. Nature 585, 217–220 (2020)

    2020

  80. [88]

    Wang, T. et al. Ab initio characterization of protein molecular dynamics with AI 2BMD. Nature 635, 1019–1027 (2024)

    2024

  81. [89]

    S., Simine, L

    Liu, Y., Madanchi, A., Anker, A. S., Simine, L. & Deringer, V. L. The amorphous state as a frontier in computational materials design. Nat. Rev. Mater., DOI: 10.1038/s41578-024- 00754-2 (2024)

    2024 doi

  82. [90]

    Miret, S., Lee, K. L. K., Gonzales, C., Mannan, S. & Krishnan, N. M. A. Energy & f orce regression on DFT t rajectories is not enough for universal machine learning interatomic potentials. Preprint at arXiv:2502.03660 (2025)

    2025 arXiv

  83. [91]

    https://pacemaker.readthedocs.io/en/latest/pacemaker/faq

    Frequently asked questions (FAQ). https://pacemaker.readthedocs.io/en/latest/pacemaker/faq

  84. [93]

    Kühne, T. D. et al. CP2K: An electronic structure and molecular dynamics software pack- age - Quickstep: Efficient and accurate electronic structure calculations. J. Chem. Phys. 152, 194103 (2020)

    2020

  85. [94]

    & Hutter, J

    Goedecker, S., Teter, M. & Hutter, J. Separable dual -space Gaussian pseudopotentials. Phys. Rev. B 54, 1703 (1996)

    1996

  86. [95]

    & Hafner, J

    Kresse, G. & Hafner, J. Ab initio molecular dynamics for liquid metals. Phys. Rev. B 47, 558–561 (1993)

    1993

  87. [96]

    & Furthmüller, J

    Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total- energy calcula- tions using a plane-wave basis set. Phys. Rev. B 54, 11169–11186 (1996)

    1996

  88. [97]

    Blöchl, P. E. Projector augmented-wave method. Phys. Rev. B 50, 17953–17979 (1994)

    1994

  89. [98]

    & Joubert, D

    Kresse, G. & Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B 59, 1758 (1999). 30

    1999

  90. [99]

    GST-ACE-24

    Thompson, A. P. et a l. LAMMPS - a flexible simulation tool for particle -based materials modeling at the atomic, meso, and continuum scales. Comput. Phys. Commun. 271, 108171 (2022). S1 Supplementary Information for Full-cycle device-scale simulations of memory materials with...

    2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.