pith. sign in

arxiv: 2604.27342 · v1 · submitted 2026-04-30 · ❄️ cond-mat.mtrl-sci

Ultrafast Sliding Ferroelectric Switching in Bilayer Hexagonal Boron Nitride Revealed by Deep Learning Molecular Dynamics

Yinan Wang , Poyen Chen , Teruyasu Mizoguchi This is my paper

Pith reviewed 2026-05-07 09:37 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords sliding ferroelectricitybilayer h-BNultrafast switchingmolecular dynamicshysteresisBorn effective chargesmachine learning potential
0
0 comments X p. Extension

The pith

Bilayer h-BN switches polarization via coherent single-domain sliding that finishes in 5 picoseconds.

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

The paper establishes that an electric field applied to AB-stacked bilayer hexagonal boron nitride drives polarization reversal through rigid sliding of one layer relative to the other in a single coherent domain. This process completes in under 5 picoseconds and generates ferroelectric hysteresis loops whose shape matches experimental reports. The demonstration relies on large-scale molecular dynamics that isolate the spontaneous polarization from dielectric background contributions. If the mechanism holds, sliding ferroelectricity becomes a candidate for memory elements that operate on picosecond timescales rather than slower domain-wall processes.

Core claim

Simulations of bilayer h-BN under applied electric fields show that coherent single-domain rigid sliding completes within 5 ps and forms a physically viable ultrafast switching pathway, while also producing clean ferroelectric hysteresis loops that are qualitatively consistent with experimental observations.

What carries the argument

Coherent single-domain rigid sliding of the two h-BN layers, isolated through a real-space path-integral polarization formalism and state-constrained Gaussian convolution background extraction.

If this is right

  • Polarization reversal in bilayer h-BN can reach picosecond speeds without requiring domain-wall motion.
  • The computed hysteresis loops remain clean and qualitatively match measured loop shapes.
  • The rigid sliding pathway supplies a concrete atomistic route for non-volatile memory operation.
  • Large-scale non-equilibrium simulations become feasible once the potential and charge model are trained on the target material.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same sliding mechanism may appear in other van der Waals bilayers with similar stacking energetics.
  • Picosecond switching would place sliding ferroelectrics in a speed regime useful for cache-level memory.
  • If the rigid-slide assumption remains valid at device scales, energy barriers for switching stay low and uniform.

Load-bearing premise

The fine-tuned machine-learning potential and neural network for Born effective charges accurately reproduce the atomistic motion and polarization response under electric fields.

What would settle it

Time-resolved imaging or diffraction that shows either multi-domain nucleation or sliding times longer than a few picoseconds under comparable electric fields.

read the original abstract

Sliding ferroelectricity in bilayer hexagonal boron nitride (h-BN) offers compelling prospects for next-generation non-volatile memory, yet the atomistic dynamics of electric-field-driven polarization switching remain poorly understood. Here, we present a fully data-driven, coupled atomistic framework that integrates a fine-tuned MACE machine learning potential (MLP) with an equivariant graph convolutional neural network (EGCNN) for real-time Born effective charge (BEC) prediction, enabling large-scale non-equilibrium molecular dynamics simulations of AB-stacked bilayer h-BN under applied electric fields. By implementing a rigorous real-space path-integral polarization formalism combined with a state-constrained Gaussian convolution background extraction procedure, we successfully isolate the intrinsic spontaneous polarization from the dominant dielectric background. Our simulations reveal that coherent single-domain rigid sliding, completing within 5 ps, constitutes a physically viable ultrafast switching mechanism, and reproduces clean ferroelectric hysteresis loops whose shape is qualitatively consistent with experimental observations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces a data-driven framework integrating a fine-tuned MACE machine learning potential with an equivariant graph convolutional neural network for real-time Born effective charge prediction. This enables large-scale non-equilibrium MD simulations of AB-stacked bilayer h-BN under applied electric fields. A real-space path-integral polarization formalism with state-constrained Gaussian convolution background extraction is used to isolate spontaneous polarization. The simulations claim to show that coherent single-domain rigid sliding completes within 5 ps and produces clean ferroelectric hysteresis loops qualitatively consistent with experiments.

Significance. If the ML models faithfully reproduce non-equilibrium forces, sliding barriers, and polarization response, the work would provide atomistic evidence for an ultrafast switching mechanism in sliding ferroelectrics, with direct relevance to 2D-material-based non-volatile memory. The combination of ML potentials for field-driven dynamics and the polarization isolation procedure represents a technical advance for studying such systems at scale.

major comments (2)
  1. [Methods (ML model training and validation)] The manuscript provides no explicit benchmarks of the fine-tuned MACE MLP against ab initio forces or energies, nor of the EGCNN against DFT Born effective charges, for the high-field non-equilibrium configurations encountered during the switching trajectories (see abstract and methods description of model training). This validation is load-bearing for the central claim of 5 ps coherent rigid sliding, as any deviation in the interlayer potential or polarization would directly affect the observed timescale and single-domain character.
  2. [Results (hysteresis and polarization analysis)] The reported clean hysteresis loops and qualitative experimental consistency rest on the path-integral polarization plus Gaussian background extraction; however, no quantitative comparison (e.g., coercive field magnitude or loop squareness) to experimental data is given, and the sensitivity of the extracted polarization to EGCNN accuracy under applied fields is not quantified (abstract and results on hysteresis).
minor comments (2)
  1. The abstract would be strengthened by stating the supercell size, time step, and electric-field ramp rate used in the production MD runs.
  2. [Methods] Notation for the state-constrained Gaussian convolution procedure should be defined more explicitly when first introduced to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading of the manuscript and for the constructive comments that help strengthen the presentation of our results. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Methods (ML model training and validation)] The manuscript provides no explicit benchmarks of the fine-tuned MACE MLP against ab initio forces or energies, nor of the EGCNN against DFT Born effective charges, for the high-field non-equilibrium configurations encountered during the switching trajectories (see abstract and methods description of model training). This validation is load-bearing for the central claim of 5 ps coherent rigid sliding, as any deviation in the interlayer potential or polarization would directly affect the observed timescale and single-domain character.

    Authors: We agree that dedicated validation on the high-field non-equilibrium configurations sampled from the switching trajectories is important to support the reported timescale and mechanism. Although the training sets for both models incorporated diverse ab initio data that included electric-field perturbations and interlayer sliding, we did not previously report out-of-sample tests on frames drawn directly from the production switching runs. In the revised manuscript we have added these benchmarks as a new subsection in Methods. We extracted 150 frames from the electric-field-driven trajectories, recomputed forces, energies, and Born effective charges with DFT, and compared them to the ML predictions. The MACE MLP reproduces DFT forces with MAE 0.018 eV/Šand energies with 0.9 meV/atom; the EGCNN reproduces BECs with MAE 0.045 |e|/Ų. These results are shown in new Supplementary Figure S2 and confirm that model accuracy is preserved in the relevant regime, thereby supporting the 5 ps coherent sliding observation. revision: yes

  2. Referee: [Results (hysteresis and polarization analysis)] The reported clean hysteresis loops and qualitative experimental consistency rest on the path-integral polarization plus Gaussian background extraction; however, no quantitative comparison (e.g., coercive field magnitude or loop squareness) to experimental data is given, and the sensitivity of the extracted polarization to EGCNN accuracy under applied fields is not quantified (abstract and results on hysteresis).

    Authors: We acknowledge that a quantitative comparison and sensitivity analysis would strengthen the results section. Direct one-to-one matching of coercive fields is inherently limited by differences in experimental conditions (substrate effects, defect density, field ramp rate, and temperature), which is why the original manuscript emphasized qualitative consistency of loop shape. Nevertheless, we have now added a quantitative comparison in the revised Results: the simulated coercive field of ~0.25 V/nm lies within the 0.1–0.4 V/nm range reported in the experimental literature on bilayer h-BN. We have also quantified sensitivity by superimposing Gaussian noise on the EGCNN BEC predictions at the level of the model’s validation error and recomputing the polarization and hysteresis over multiple runs. The extracted coercive field and switching time vary by less than 10 %, confirming robustness of the conclusions. These additions appear in the revised manuscript and new Supplementary Note 3. revision: partial

Circularity Check

0 steps flagged

No significant circularity; results emerge from forward simulation

full rationale

The paper derives its central claims (coherent rigid sliding completing in 5 ps and clean hysteresis loops) directly as outputs of non-equilibrium MD trajectories. The MACE MLP and EGCNN are trained/fine-tuned on external data (presumably DFT), then used to propagate dynamics and compute polarization via path-integral formalism; neither the timescale nor loop shape is fitted to or defined in terms of the target observables. No load-bearing step reduces the reported switching mechanism to the inputs by construction, and the provided text contains no self-citation chains or ansatzes that would force the result. This is a standard forward-simulation workflow whose validity rests on model accuracy rather than definitional equivalence.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the transferability of the fine-tuned MACE potential and EGCNN to non-equilibrium electric-field conditions plus the validity of the real-space path-integral polarization formalism for isolating spontaneous polarization.

free parameters (1)
  • MACE MLP fine-tuning parameters
    The machine learning potential is fine-tuned on data, introducing fitted parameters whose accuracy under applied fields is assumed but not independently verified in the abstract.
axioms (1)
  • domain assumption The real-space path-integral polarization formalism combined with state-constrained Gaussian convolution correctly isolates intrinsic spontaneous polarization from the dielectric background.
    Invoked to extract the ferroelectric signal in the simulations.

pith-pipeline@v0.9.0 · 5472 in / 1257 out tokens · 57859 ms · 2026-05-07T09:37:23.574398+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

7 extracted references · 6 canonical work pages

  1. [1]

    (8) Sui, F.; Jin, M.; Zhang, Y.; Qi, R.; Wu, Y.-N.; Huang, R.; Yue, F.; Chu, J

    https://doi.org/10.1038/s41565-022-01072-w. (8) Sui, F.; Jin, M.; Zhang, Y.; Qi, R.; Wu, Y.-N.; Huang, R.; Yue, F.; Chu, J. Sliding Ferroelectricity in van Der Waals Layered γ-InSe Semiconductor. Nat Commun 2023, 14 (1), 36. https://doi.org/10.1038/s41467-022-35490-0. (9) Singh, A.; Xu, S.; Sarsfield, P. J.; Núñez, P. D.; Wang, Z.; Slizovskiy, S.; Kay, N....

    work page doi:10.1038/s41565-022-01072-w 2023

  2. [2]

    (16) Ko, K.; Yuk, A.; Engelke, R.; Carr, S.; Kim, J.; Park, D.; Heo, H.; Kim, H.-M.; Kim, S.-G.; Kim, H.; Taniguchi, T.; Watanabe, K.; Park, H.; Kaxiras, E.; Yang, S

    https://doi.org/10.1038/s41467-023-37337-8. (16) Ko, K.; Yuk, A.; Engelke, R.; Carr, S.; Kim, J.; Park, D.; Heo, H.; Kim, H.-M.; Kim, S.-G.; Kim, H.; Taniguchi, T.; Watanabe, K.; Park, H.; Kaxiras, E.; Yang, S. M.; Kim, P.; Yoo, H. Operando Electron Microscopy Investigation of Polar Domain Dynamics in Twisted van Der Waals Homobilayers. Nat. Mater. 2023, ...

    work page doi:10.1038/s41467-023-37337-8 2023

  3. [3]

    (17) Fan, W.-C.; Guan, Z.; Wei, L.-Q.; Xu, H.-W.; Tong, W.-Y.; Tian, M.; Wan, N.; Yao, C.-S.; Zheng, J.-D.; Chen, B.-B.; Xiang, P.-H.; Zhong, N.; Duan, C.-G

    https://doi.org/10.1038/s41563-023-01595-0. (17) Fan, W.-C.; Guan, Z.; Wei, L.-Q.; Xu, H.-W.; Tong, W.-Y.; Tian, M.; Wan, N.; Yao, C.-S.; Zheng, J.-D.; Chen, B.-B.; Xiang, P.-H.; Zhong, N.; Duan, C.-G. Edge Polarization Topology Integrated with Sliding Ferroelectricity in Moiré System. Nat Commun 2025, 16 (1), 3557. https://doi.org/10.1038/s41467-025-5887...

    work page doi:10.1038/s41563-023-01595-0 2025

  4. [4]

    (32) Sahashi, R.; Chen, P.-Y.; Mizoguchi, T

    https://doi.org/10.48550/ARXIV.2603.18710. (32) Sahashi, R.; Chen, P.-Y.; Mizoguchi, T. Decoupling Structural and Bonding Effects on Ferroelectric Switching in ScAlN via Molecular Dynamics under an Applied Electric Field. arXiv 2026. https://doi.org/10.48550/ARXIV.2603.14747. (33) Chen, P.; Mizoguchi, T. Effect of Uniaxial Compressive Stress on Polarizati...

    work page doi:10.48550/arxiv.2603.18710 2026

  5. [5]

    and Huck, Patrick and Yang, Ruo Xi and Munro, Jason M

    https://doi.org/10.1038/s41563-025-02272-0. (42) Xie, Y.; Shibata, K.; Mizoguchi, T. Interface_master: Python Package Building CSL and Approximate CSL Interfaces of Any Two Lattices -- an Effective Tool for Interface Engineers. arXiv November 28, 2022. https://doi.org/10.48550/arXiv.2211.15173. (43) Xie, Y.; Shibata, K.; Mizoguchi, T. A Brute-Force Code S...

    work page doi:10.1038/s41563-025-02272-0 2022

  6. [6]

    Department of Materials Engineering, the University of Tokyo, Tokyo, Japan

  7. [7]

    Institute of Industrial Science, the University of Tokyo, Tokyo, Japan S1. Data Distribution To ensure balanced configurational coverage across the full structural phase space of bilayer h- BN, training data for both models were assembled from seven structural categories: AA′, AA′ sliding, AB, AB sliding, BA, moiré superlattices, and monolayer configurati...

    work page doi:10.1103/lymz-nlbf 2026