REVIEW 3 major objections 5 minor 1 cited by
Accurate Modeling of Interfacial Thermal Transport in van der Waals Heterostructures via Hybrid Machine Learning and Registry-Dependent Potentials
T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read A hybrid machine-learning plus interlayer potential matches DFT for TMD heat transport.
desk verdict Solid six-TMD benchmark for a hybrid NEP+ILP approach, but the abstract oversells the bulk modulus accuracy and the reproducibility artifacts are incomplete. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the total energy $E_{\text{tot}}^{\text{NEP+ILP}}$, which splits pairwise interactions between layers: the neuroevolution potential handles intralayer chemistry through radial Chebyshev descriptors and angular spherical-harmonic descriptors up to degree $l=4$ with a 5 Å cutoff, while the registry-dependent interlayer potential handles van der Waals interactions with an anisotropic repulsive term that depends on lateral registry distances $\rho_{ij}$ relative to the local layer normals and a damped $C_6/r^6$ attraction out to a 16 Å cutoff. That registry dependence is what lets the model distinguish AA, AB, and saddle stacking, so it carries the interlayer sliding energy surface and the moiré reconstruction; the NEP carries phonon and anharmonic properties.
What would settle it
Recompute the ILP reference binding-energy curves and sliding potential energy surfaces for MoTe2 and WTe2 with HSE+MBD-NL and rerun the Section 3 benchmarks; if the equilibrium interlayer distances or sliding corrugations shift by more than the spread quoted in Supporting Information Section 2, the reported c0 values, bulk moduli, and cross-plane conductivities for these two compounds are biased. An experimental check is X-ray compression data: measured c0(P) of MoTe2 and WTe2 should match the predicted 14.29 Å and 14.34 Å spacings and the Murnaghan moduli of 24.5 and 23.8 GPa.
Extended reading notes
Core claim
The central discovery is that decoupling intralayer and interlayer interactions—training the NEP on monolayer configurations only, and parameterizing the ILP on bilayer binding-energy curves and sliding potential energy surfaces—preserves accuracy while cutting training data. With this split, the NEP+ILP force field reproduces room-temperature in-plane conductivities (e.g., 150.4 W/m/K for monolayer MoS2 and 216.2 W/m/K for WS2), cross-plane bulk values that align with measurements (4.1 and 6.1 W/m/K, respectively), phonon dispersions close to PBE+MBD, and the triangular and hexagonal moiré reconstruction patterns seen in twisted TMD experiments. For the telluride systems the ILP parameters rest on PBE+MBD-NL reference data because HSE+MBD-NL is too costly; the paper argues the difference is small and benchmarks it in the supporting information. A head-to-head comparison for MoS2 against the NEP-D3 alternative shows the hybrid method captures interlayer binding and sliding energy surfaces with near-HSE+MBD-NL fidelity, while NEP-D3 overestimates them.
Load-bearing premise
The load-bearing premise is that PBE+MBD-NL density functional theory can stand in for the more expensive HSE+MBD-NL when fitting the interlayer potential of MoTe2 and WTe2, so if the two functionals diverge more than the supporting comparison indicates, every computed property for those two materials inherits that bias.
Editorial extensions
If this is right
- Researchers can simulate micron-scale twisted TMD stacks with DFT-level forces, allowing direct prediction of how moiré reconstruction alters interfacial thermal conductance with twist angle.
- The decoupled training strategy means adding a new TMD pair requires only monolayer DFT frames plus a few bilayer energy curves, drastically lowering the cost of building accurate potentials across the MX2 family.
- Cross-plane conductivity values near experimental data imply the interlayer potential is reliable for through-plane phonon transport, so the method can be used to design van der Waals thermal metamaterials by stacking and twisting layers.
- The MoS2 comparison indicates NEP+ILP outperforms NEP-D3 in sliding-energy fidelity, making it an appropriate base for friction and superlubricity simulations of TMD interfaces.
Reading between the lines
- If the PBE-vs-HSE assumption holds, the same workflow should extend cleanly to other telluride or heavier chalcogenide pairs; if it fails, the MoTe2 and WTe2 predictions offer a clear way to detect the bias by measuring interlayer spacing under pressure.
- The framework's registry sensitivity invites a connected prediction not made in the paper: the cross-plane thermal conductance of a twisted TMD bilayer should be tunable by sliding one layer at fixed twist angle, since the sliding potential corrugation is captured accurately.
- The NEP-D3 comparison suggests a general design principle for dispersion-corrected machine-learned potentials: including sliding configurations at the true equilibrium interlayer distance in the training set is necessary for transferability.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript extends a hybrid machine-learning/registry-dependent interlayer potential framework (NEP+ILP) to six TMD homojunctions (MoS2, MoSe2, MoTe2, WS2, WSe2, WTe2). The intralayer interactions are described by a neuroevolution potential (NEP) trained on monolayer DFT data, and the interlayer interactions by an anisotropic registry-dependent ILP fitted to binding-energy curves and sliding potential energy surfaces. The authors benchmark the combined potential against lattice constants, bulk modulus, moiré reconstruction, phonon spectra, and in-plane and cross-plane thermal conductivities, and also compare with an alternative NEP-D3 approach for MoS2. The central claim, stated in the abstract and conclusions, is that the NEP+ILP framework achieves near-quantum-mechanical accuracy for key TMD properties while remaining scalable to large moiré superlattices.
Significance. If the hybrid framework delivers DFT-level accuracy for interlayer registry effects and thermal transport at classical MD cost, it would be a practically valuable tool for studying TMD heterostructures, which are difficult to treat by direct DFT at moiré length scales. The paper includes useful strengths: the potentials and code are made available, the cross-plane thermal conductivities agree well with available experiments, and the direct comparison with NEP-D3 provides a concrete assessment of an alternative approach. However, the paper's own benchmark data show that the bulk modulus is underestimated by 30–59% relative to experiments for five of the six systems, which directly contradicts the unqualified 'accurately predicts bulk modulus' claim in the abstract. This is a load-bearing discrepancy because bulk modulus is listed among the headline validated properties, and a too-soft interlayer response under compression may affect pressure-dependent properties and moiré reconstruction energetics.
major comments (3)
- [Section 3.2 and Fig. 2] The reported Murnaghan bulk moduli are 35.1, 29.7, 24.5, 25.9, 29.8, and 23.8 GPa for MoS2, MoSe2, MoTe2, WS2, WSe2, and WTe2, and the manuscript states that for MoS2, MoSe2, MoTe2, WS2, and WSe2 these underestimate experimental results by about 30.5%, 44.8%, 42.3%, 58.2%, and 58.6%, respectively. This is more than a minor deviation, and it contradicts the abstract's statement that the framework 'accurately predicts' the bulk modulus. Because the bulk modulus is one of the headline properties in the central claim, the overstatement should be corrected and, more importantly, the origin of the systematic softness of the interlayer potential under hydrostatic compression should be discussed or mitigated; if it cannot be mitigated, the claims must be limited to properties for which the demonstrated accuracy is supported.
- [Section 2.3] The ILP parameters for MoTe2 and WTe2 are fitted to PBE+MBD-NL reference data because HSE+MBD-NL is deemed computationally prohibitive, and the manuscript justifies this by stating that the disparity between PBE+MBD-NL and HSE+MBD-NL is 'small', with the quantitative comparison deferred to Section 2 of the Supporting Information. Since the Te-based systems comprise one third of the benchmark set, this substitution is load-bearing: any systematic difference between PBE and HSE interlayer binding or sliding energies would directly bias the lattice constants, bulk moduli, phonon spectra, and thermal conductivities reported for MoTe2 and WTe2. The quantitative evidence for this equivalence should be presented in the main text or the argument should be made convincingly in the Supporting Information; the current one-sentence assertion is not sufficient.
- [Abstract and Section 5] The abstract claims 'near quantum mechanical accuracy' and 'accurately predicting key properties such as lattice constants, bulk modulus, moiré reconstruction, phonon spectra, and thermal conductivities', while the conclusions say the framework yields 'bulk modulus values and in-plane thermal conductivity in good agreement' with DFT and 'cross-plane thermal conductivity aligned with experimental results'. Given the measured bulk-modulus errors of 30–59% and the in-plane thermal conductivities that are systematically higher than most experiments (though consistent with BTE predictions), the phrase 'good agreement' is too strong. The claims should be recalibrated to state exactly which properties agree quantitatively and which agree only semi-quantitatively or with known systematic deviations.
minor comments (5)
- [Throughout] There are several typographical errors, including 'respecitvley' in Section 3.3, 'diomands' in Figure 8, and 'bashed' in the Table 1 caption; these should be corrected.
- [Equation (5) and surrounding text] The notation in Equation (5) is confusing: the neighbor sets Ω_k and Ω_l are introduced, but the text then refers to 'atom i in layer k and atom i in layer l' with inconsistent subscripts. Please clarify the index convention.
- [Section 3.5] The monolayer thicknesses used for converting 2D to 3D thermal conductivity are defined (6.2, 6.6, 7.2 Å), but the choice of these values and their sensitivity to the reported thermal conductivities are not discussed; a brief justification or a reference for the effective thickness convention would be helpful.
- [Figure captions] In Figure 2, the caption states that 'Reported experimental values and our MD simulation results are presented as blue and red circles', but the figure also includes DFT data; please clarify the symbol and color coding for all three data types.
- [Section 3.1] The deviations of the lattice constants are reported as '0.01, 0.02, 0.02, 0.01, 0.02, and 0.02 Å from experiments and DFT data', but it is unclear whether these are maximum deviations, mean absolute deviations, or deviations from a specific reference; please specify.
Circularity Check
Minor circularity in the MoS2 interlayer-energy validation; the headline thermal-transport predictions are not fitted values.
-
fitted input called prediction
[Section 4, Fig. 8(d-i) and caption]
"In contrast, NEP+ILP demonstrates good agreement with HSE+MBD-NL in interlayer energy calculations, further validating its suitability for modeling interfacial mechanics in TMD materials. ... All the HSE+MBD-NL results are extracted from Ref. 24."
For MoS2, the ILP parameters were parameterized in the authors' earlier Ref. 24 against HSE+MBD-NL binding-energy curves and sliding PESs. The present paper then compares NEP+ILP with those same HSE+MBD-NL data and calls the agreement a validation. This is a training-set reproduction: the ILP was fitted to that exact reference data, so the good agreement is by construction, not an independent check. The step is secondary because it supports the NEP+ILP versus NEP-D3 comparison for interfacial energies rather than the central thermal-conductivity predictions, which are generated by HNEMD from potentials fitted to non-thermal DFT data.
full rationale
Most of the derivation chain is self-contained. The NEP is trained on monolayer DFT energies/forces/virials, and the ILP is trained on DFT binding-energy curves and sliding PESs; the headline in-plane and cross-plane thermal conductivities are then produced by HNEMD simulations and are not fit to any thermal data. Lattice constants, phonon spectra, and bulk moduli are simulated outputs benchmarked against external DFT/experimental references. The only identifiable circular element is the MoS2 interlayer-energy comparison in Section 4/Fig. 8: the ILP used there was fitted to the very HSE+MBD-NL binding/sliding data shown as the reference, making that agreement a consistency check rather than an independent validation. This does not affect the heat-transport predictions. The large bulk-modulus underestimates reported in Sec. 3.2 are an accuracy and overclaim issue, not a circularity, because the P-V curves are simulated outputs and no fitted parameter is relabeled as a prediction.
Assumptions & free parameters
free parameters (4)
- ILP parameters (epsilon_ij, C_ij, beta_ij, gamma_ij, alpha_ij, C6_ij, r_eff_ij, d_ij, s_R_ij) =
Not tabulated in main text
- NEP neural-network weights and biases =
Not provided in manuscript
- Effective monolayer thickness =
6.2, 6.6, 7.2 Å depending on chalcogen
- NEP descriptor cutoffs and hyperparameters =
rc^R = rc^A = 5 Å; 9 radial functions, 13 Chebyshev polynomials
assumptions (6)
- domain assumption Intralayer and interlayer interactions are additive and separable (Eq. 1).
- domain assumption The ILP functional form (Eqs. 5-8) captures the registry-dependent vdW interaction.
- domain assumption HSE+MBD-NL DFT is an accurate ground truth for S/Se TMD interlayer energies.
- ad hoc to paper PBE+MBD-NL is an accurate substitute for HSE+MBD-NL for MoTe2 and WTe2.
- domain assumption The NEP model trained on monolayer configurations transfers unchanged to multilayer and bulk systems.
- domain assumption The effective thickness values used for kappa conversion are physically appropriate.
Cite this review
Pith. "Pith review of Accurate Modeling of Interfacial Thermal Transport in van der Waals Heterostructures via Hybrid Machine Learning and Registry-Dependent Potentials." pith.science (2026). https://pith.science/paper/KE5KYBVX
@misc{pith2026250500376,
author = {Pith},
title = {Pith review of: Accurate Modeling of Interfacial Thermal Transport in van der Waals Heterostructures via Hybrid Machine Learning and Registry-Dependent Potentials},
year = {2026},
howpublished = {\url{https://pith.science/paper/KE5KYBVX}},
note = {Machine review of arXiv:2505.00376}
}
read the original abstract
Two-dimensional transition metal dichalcogenides (TMDs) exhibit remarkable thermal anisotropy due to their strong intralayer covalent bonding and weak interlayer van der Waals (vdW) interactions. However, accurately modeling their thermal transport properties remains a significant challenge, primarily due to the computational limitations of density functional theory (DFT) and the inaccuracies of classical force fields in non-equilibrium regimes. To address this, we use a recently developed hybrid computational framework that combines machine learning potential (MLP) for intralayer interactions with registry-dependent interlayer potential (ILP) for anisotropic vdW interlayer interaction, achieving near quantum mechanical accuracy. This approach demonstrates exceptional agreement with DFT calculations and experimental data for TMD systems, accurately predicting key properties such as lattice constants, bulk modulus, moir\'e reconstruction, phonon spectra, and thermal conductivities. The scalability of this method enables accurate simulations of TMD heterostructures with large-scale moir\'e superlattices, making it a transformative tool for the design of TMD-based thermal metamaterials and devices, bridging the gap between accuracy and computational efficiency.
Forward citations
Cited by 1 Pith paper
-
NEP89: Universal neuroevolution potential for inorganic and organic materials across 89 elements
NEP89 is a universal neuroevolution potential that combines near-DFT accuracy with empirical-potential speed for atomistic simulations of inorganic and organic materials across 89 elements.
Reference graph
Works this paper leans on
-
[2]
ACS Appl. Mater. Interfaces 13, 13063- 13071 (2021). (107) Taube, A., Judek, J., Łapińska, A., Zdrojek, M. Temperature-dependent thermal properties of supported MoS 2 monolayers. ACS Appl. Mater. Interfaces 7, 5061-5065 (2015). (108) Yang, X., Zheng, X., Liu, Q., Zhang, T., Bai, Y ., Yang, Z., Chen, H., Liu, M. Experimental study on thermal conductivity a...
work page 2021
-
[3]
ACS Appl. Mater. Interfaces 12, 28306-28312 (2020). (109) Zhang, X., Sun, D., Li, Y ., Lee, G.-H., Cui, X., Chenet, D., You, Y ., Heinz, T. F., Hone, J. C. Measurement of lateral and interfacial thermal conductivity of s ingle- and bilayer MoS2 and MoSe2 using refined optothermal raman technique. ACS Appl. Mater. Interfaces 7, 25923-25929 (2015). (110) Ya...
work page 2020
-
[4]
Phys. Chem. Chem. Phys. 24, 13860-13868 (2022). (131) Shen, J., Han, D., Zhang, B., Cao, R., Liu, Y ., Zheng, S., Li, H., Jiang, Y ., Xue, Y ., Xue, M. First-principles study on phonon transport properties of MoTe 2 and WTe2 monolayers in different phases. Phys. E 145, 115509 (2023). (132) Shafique, A., Shin, Y .- H. Strain engineering of phonon thermal t...
work page 2022
-
[5]
Phys. Rev. B 108, 115439 (2023). (135) Mobaraki, A., Sevik, C., Yapicioglu, H., Çakır, D., Gülseren, O. Temperature -dependent phonon spectrum of transition metal dichalcogenides calculated from the spectral energy density: Lattice thermal conductivity as an application. Phys. Rev. B 100, 035402 (2019). (136) Xu, K., Gabourie, A. J., Hashemi, A., Fan, Z.,...
work page 2023
-
[6]
Phys. Rev. Mater. 2, 114008 (2018). (138) Peng, B., Zhang, H., Shao, H., Xu, Y ., Zhang, X., Zhu, H. Thermal conductivity of monolayer MoS2, MoSe2, and WS2: Interplay of mass effect, interatomic bonding and anharmonicity. RSC Adv. 6, 5767-5773 (2016). (139) Zulfiqar, M., Zhao, Y ., Li, G., Li, Z., Ni, J. Intrinsic Thermal conducti vities of monolayer 29 t...
work page 2018
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.