pith. sign in

arxiv: 2605.22601 · v1 · pith:W7WZTPGBnew · submitted 2026-05-21 · ❄️ cond-mat.mtrl-sci

Upscaling DFT-trained machine-learning interatomic potential toward Quantum Monte Carlo accuracy: Sulfur-vacancy migration in monolayer MoS₂ as a testbed

Adam Hlo\v{z}n\'y , J\'an Brndiar , Ye Luo , Ivan \v{S}tich This is my paper

Pith reviewed 2026-05-22 03:59 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords machine learning interatomic potentialquantum Monte Carlomulti-fidelity trainingMoS2sulfur vacancydefect migrationMACE model
0
0 comments X

The pith

A machine learning interatomic potential reaches near quantum Monte Carlo accuracy by fine-tuning readout layers with QMC energies and DFT forces

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

The paper develops a training procedure that brings machine learning interatomic potentials close to the accuracy of quantum Monte Carlo calculations for material defects. It pairs expensive QMC energies with cheaper processed DFT forces to update only the final readout layers of a pre-trained equivariant model. The method is tested on the migration of single and multiple sulfur vacancies in monolayer molybdenum disulfide. A sympathetic reader would care because direct QMC calculations remain too costly for large numbers of atoms or long molecular dynamics runs, so this approach opens the door to high-accuracy simulations at practical scales.

Core claim

We designed a procedure to train a machine learning interatomic potential (MLIP) at benchmark-quality quantum Monte Carlo (QMC) accuracy. To avoid the complexities of high-quality atomic force determination with the stochastic QMC methods, we use a multi-fidelity approach wherein high-level QMC energies are used alongside suitably processed low-level DFT atomic forces to train a QMC fine-tuned MLIP which significantly improves both the energetics and atomic forces over the baseline DFT-based MLIP. Fine-tuning is only applied to the readout layers of an equivariant message-passing MACE MLIP. We used sulfur mono- and multiple vacancies in monolayer MoS2 as a testbed and demonstrate a near QMC

What carries the argument

Multi-fidelity fine-tuning applied only to the readout layers of an equivariant message-passing MACE MLIP, using QMC energies together with processed DFT forces

If this is right

  • Large-scale simulations involving large numbers of atoms become feasible at near QMC quality.
  • Molecular dynamics runs with many configurations can reach near QMC accuracy.
  • Energy and free energy migration barriers for mono- and multiple S-vacancy defects can be computed with improved accuracy.
  • A fairly limited dataset of QMC energies is enough to produce significant gains over a baseline DFT MLIP.
  • The resulting model maintains near QMC accuracy across both in-domain and out-of-domain tests.

Where Pith is reading between the lines

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

  • The same readout-layer update could be tested on other 2D materials or vacancy types once small QMC reference sets exist.
  • Preserving force consistency from DFT while upgrading energetics with QMC may generalize to other multi-fidelity material modeling tasks.
  • The approach invites experiments that gradually reduce the QMC dataset size to map the minimum data needed for target accuracy.

Load-bearing premise

Suitably processed low-level DFT atomic forces combined with a fairly limited dataset of QMC energies are sufficient to fine-tune only the readout layers of an equivariant message-passing MACE MLIP and produce transferable near-QMC accuracy without overfitting or loss of force consistency.

What would settle it

Perform fresh QMC calculations on an out-of-domain configuration such as a larger supercell or different vacancy arrangement in MoS2 and check whether the fine-tuned MLIP predictions for energy and forces deviate beyond the reported near-QMC error level.

Figures

Figures reproduced from arXiv: 2605.22601 by Adam Hlo\v{z}n\'y, Ivan \v{S}tich, J\'an Brndiar, Ye Luo.

Figure 1
Figure 1. Figure 1: FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
read the original abstract

We designed a procedure to train a machine learning interatomic potential (MLIP) at benchmark-quality quantum Monte Carlo (QMC) accuracy. To avoid the complexities of high-quality atomic force determination with the stochastic QMC methods, we use a multi-fidelity approach wherein high-level QMC energies are used alongside suitably processed low-level DFT atomic forces to train a QMC fine-tuned MLIP which significantly improves both the energetics and atomic forces over the baseline DFT-based MLIP. Fine-tuning is only applied to the readout layers of an equivariant message-passing MACE MLIP. We used sulfur mono- and multiple vacancies in monolayer MoS$_2$ as a testbed and demonstrate a near QMC accuracy of the model in a number of in- and out-of-domain tests. We show that a fairly limited dataset of QMC energies suffice to significantly improve the baseline DFT MLIP. The accuracy of our approach is demonstrated on energy and free energy migration barriers of mono- and multiple S-vacancy defects. The results open the window to large-scale near QMC quality simulations with large numbers of atoms and/or molecular dynamics configurations which would not be possible by a direct brute-force application of QMC methods.

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 manuscript describes a multi-fidelity procedure for training an equivariant message-passing MACE MLIP on monolayer MoS2 with S-vacancies. High-level QMC energies are combined with processed low-level DFT forces to fine-tune only the readout layers of a DFT-pretrained model, with the goal of reaching near-QMC accuracy on energy and free-energy migration barriers for mono- and multiple-vacancy defects. The approach is tested on both in- and out-of-domain configurations and is presented as enabling large-scale simulations inaccessible to direct QMC.

Significance. If the central claims are substantiated, the work offers a practical route to near-QMC-quality molecular dynamics on systems with hundreds of atoms or long trajectories, which is significant for defect migration studies in 2D materials. Credit is due for the independent use of external QMC and DFT targets, the avoidance of direct QMC force calculations, and the choice of a concrete, falsifiable testbed (S-vacancy barriers).

major comments (2)
  1. [§4] §4 (Fine-tuning and force consistency): the central claim that readout-only updates on a limited QMC energy set preserve transferable force accuracy is load-bearing for the reported barriers, yet no quantitative post-fine-tuning force MAE or consistency checks on out-of-domain vacancy configurations or MD trajectories are shown. Without these, it remains unclear whether the DFT-inherited forces remain consistent with the QMC-corrected energy surface.
  2. [§5.2] §5.2 (Validation metrics): the abstract and results summary assert 'near QMC accuracy' and that 'a fairly limited dataset of QMC energies suffice,' but no numerical error metrics (energy MAE, force MAE, or barrier deviations versus direct QMC) or exact QMC dataset sizes are reported. This absence prevents assessment of whether the improvement is statistically robust or risks overfitting the readout layers.
minor comments (2)
  1. The phrase 'suitably processed low-level DFT atomic forces' appears without a precise description of the processing pipeline (e.g., any scaling, filtering, or weighting); adding this detail in the methods would aid reproducibility.
  2. Figure captions for the barrier plots should explicitly state whether the shown energies are from the fine-tuned MLIP, baseline DFT-MLIP, or direct QMC reference.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and positive review, which highlights both the potential impact of the work and areas where the presentation can be strengthened. We address the two major comments point by point below and have revised the manuscript to incorporate the requested clarifications and additional checks.

read point-by-point responses

  1. Referee: §4 (Fine-tuning and force consistency): the central claim that readout-only updates on a limited QMC energy set preserve transferable force accuracy is load-bearing for the reported barriers, yet no quantitative post-fine-tuning force MAE or consistency checks on out-of-domain vacancy configurations or MD trajectories are shown. Without these, it remains unclear whether the DFT-inherited forces remain consistent with the QMC-corrected energy surface.

    Authors: We agree that explicit post-fine-tuning force metrics would provide stronger support for the central claim. Although the original manuscript demonstrates that the fine-tuned model yields migration barriers consistent with direct QMC and remains stable in MD, we acknowledge the absence of dedicated force-error quantification after readout-layer updates. In the revised manuscript we have added a new table in §4 reporting force MAE on both in-domain and out-of-domain vacancy configurations before and after fine-tuning. We have also included a consistency check in which forces from the fine-tuned model are compared against DFT forces on configurations sampled from short MD trajectories; these checks confirm that force accuracy is preserved at a level sufficient for stable dynamics and that no large inconsistencies are introduced between the corrected energies and the inherited forces. revision: yes

  2. Referee: §5.2 (Validation metrics): the abstract and results summary assert 'near QMC accuracy' and that 'a fairly limited dataset of QMC energies suffice,' but no numerical error metrics (energy MAE, force MAE, or barrier deviations versus direct QMC) or exact QMC dataset sizes are reported. This absence prevents assessment of whether the improvement is statistically robust or risks overfitting the readout layers.

    Authors: We accept that the abstract and summary sections would benefit from explicit numerical values to allow readers to judge the magnitude of improvement and the risk of overfitting. The manuscript already states the QMC dataset size in the methods and provides qualitative validation on out-of-domain structures, but we agree these could be presented more quantitatively. In the revised version we have updated the abstract and expanded §5.2 with a table that reports energy MAE, force MAE, and barrier deviations (with uncertainties) relative to direct QMC on both training and held-out sets. Cross-validation results are now shown to demonstrate that performance on out-of-domain configurations remains comparable, supporting that the limited QMC data improves accuracy without evident overfitting of the readout layers. revision: yes

Circularity Check

0 steps flagged

No significant circularity: independent QMC targets and DFT forces drive fine-tuning

full rationale

The derivation chain relies on external QMC energies as independent high-fidelity targets combined with processed DFT forces to fine-tune only the readout layers of a pre-trained MACE model. This multi-fidelity procedure does not reduce any claimed accuracy or barrier prediction to a fitted parameter by construction, nor does it invoke self-citations or uniqueness theorems that collapse the central result. In- and out-of-domain tests on vacancy migration barriers serve as external validation rather than tautological outputs, keeping the approach self-contained against the provided benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on the standard assumption that QMC energies are more accurate than DFT for the systems studied and that limited high-fidelity data can correct a pre-trained model without new physical postulates.

free parameters (1)
  • QMC dataset size
    Described only as 'fairly limited'; exact number and selection criteria not stated in abstract.
axioms (1)
  • domain assumption QMC provides benchmark-quality energies superior to DFT for the defect systems considered
    Invoked when stating that QMC energies are used to reach benchmark accuracy.

pith-pipeline@v0.9.0 · 5781 in / 1356 out tokens · 61487 ms · 2026-05-22T03:59:47.193031+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

33 extracted references · 33 canonical work pages

  1. [1]

    Behler and M

    J. Behler and M. Parrinello. Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces . Phys. Rev. Lett. , 98(14):146401, 2007

    work page 2007

  2. [2]

    A. P. Bart \'o k, M. C. Payne, R. Kondor, and G. Cs \'a nyi. Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons . Phys. Rev. Lett. , 104(13):136403, 2010

    work page 2010

  3. [3]

    J. Behler. Perspective: Machine learning potentials for atomistic simulations . J. Chem. Phys. , 145(17):170901, 2016

    work page 2016

  4. [4]

    W. M. C. Foulkes, L. Mitas, R. J. Needs, and G. Rajagopal. Quantum Monte Carlo simulations of solids . Rev. Mod. Phys. , 73(1):33--83, 2001

    work page 2001

  5. [5]

    Wines, J

    D. Wines, J. Ahn, A. Benali, P. R. C. Kent, J. T. Krogel, Y. Kwon, L. Mitas, F. A. Reboredo, B. Rubenstein, K. Saritas, H. Shin, I. S tich, and C. Ataca. Toward improved property prediction of 2D materials using many-body quantum Monte Carlo methods . Appl. Phys. Rev. , 12(3):031317, 08 2025

    work page 2025

  6. [6]

    Nakano, M

    K. Nakano, M. Casula, and G. Tenti. Efficient calculation of unbiased atomic forces in ab initio variational Monte Carlo . Phys. Rev. B , 109(20):205151, 2024

    work page 2024

  7. [7]

    van Rhijn, C

    J. van Rhijn, C. Filippi, S. De Palo, and S. Moroni. Energy Derivatives in Real-Space Diffusion Monte Carlo . J. Chem. Theory Comput. , 18(1):118--123, 2022

    work page 2022

  8. [8]

    Slootman, I

    E. Slootman, I. Poltavsky, R. Shinde, J. Cocomello, S. Moroni, A. Tkatchenko, and C. Filippi. Accurate Quantum Monte Carlo Forces for Machine-Learned Force Fields: Ethanol as a Benchmark . J. Chem. Theory Comput. , 20:6020--6027, 2024

    work page 2024

  9. [9]

    Nandi, P

    A. Nandi, P. Pandey, P. L. Houston, Ch. Qu, Q. Yu, R. Conte, A. Tkatchenko, and J. M. Bowman. Delta-Machine Learning to Elevate DFT-Based Potentials and a Force Field to the CCSD(T) Level Illustrated for Ethanol . J. Chem. Theory Comput. , 20(20):8807--8819, 2024

    work page 2024

  10. [10]

    M \'e sz \'a ros, A

    B B. M \'e sz \'a ros, A. Szab \'o , and J. Daru. Short-Range Delta-Machine Learning: A Cost-Efficient Strategy to Transfer Chemical Accuracy to Condensed Phase Systems . J. Chem. Theory Comput. , 21(11):5372--5381, 2025

    work page 2025

  11. [11]

    O'Neill, B

    N. O'Neill, B. X. Shi, W. J. Baldwin, W. C. Witt, G. Cs \'a nyi, J. D. Gale, A. Michaelides, and Ch. Schran. Towards Routine Condensed Phase Simulations with Delta-Learned Coupled Cluster Accuracy: Application to Liquid Water . J. Chem. Theory Comput. , 21(22):11710--11720, 2025

    work page 2025

  12. [12]

    and Nakano, K

    Tenti, G. and Nakano, K. and Tirelli, A.and Sorella, S. and Casula, M. Principal deuterium hugoniot via quantum monte carlo and -learning. Phys. Rev. B , 110:L041107, 2024

    work page 2024

  13. [13]

    Radova, W

    M. Radova, W. G. Stark, C. S. Allen, R. J. Maurer, and A. P. Bart \'o k. Fine-tuning foundation models of materials interatomic potentials with frozen transfer learning . npj Comput. Mat. , 11(1):237, Jul 2025

    work page 2025

  14. [14]

    Batatia, D

    I. Batatia, D. P. Kov \'a cs, G. N. C. Simm, Chr. Ortner, and G. Cs \'a nyi. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields . In Adv. Neural Inf. Process. Syst. , volume 35, 2022

    work page 2022

  15. [15]

    Messerly, S

    M. Messerly, S. Matin, A. E. A. Allen, B. Nebgen, K. Barros, J. S. Smith, N. Lubbers, and R. Messerly. Multi-fidelity learning for interatomic potentials: low-level forces and high-level energies are all you need . Mach. Learn.: Sci. Technol. , 8:035066, 2025

    work page 2025

  16. [16]

    Komsa and A

    H.-P. Komsa and A. V. Krasheninnikov. Native defects in bulk and monolayer MoS _2 from first principles . Phys. Rev. B , 91(12):125304, 2015

    work page 2015

  17. [17]

    W. Zhou, X. Zou, S. Najmaei, Z. Liu, Y. Shi, J. Kong, J. Lou, P. M. Ajayan, B. I. Yakobson, and J.-C. Idrobo. Intrinsic Structural Defects in Monolayer Molybdenum Disulfide . Nano Lett. , 13(6):2615--2622, 2013

    work page 2013

  18. [18]

    Hlo z n \'y , J

    A. Hlo z n \'y , J. Brndiar, M. Casula, and I. S tich. Structure and dynamics of sulfur vacancies in monolayer MoS _2 studied by DFT -based machine learning potentials . J. Chem. Phys. , 163(21):214118, 2025

    work page 2025

  19. [19]

    Frenkel and Smit

    D. Frenkel and Smit. B. Understanding Molecular Simulation: From Algorithms to Applications . Academic Press, Elsevier London, San Diego, Cambridge, MA, Oxford, 2007

    work page 2007

  20. [20]

    Henkelman, B

    G. Henkelman, B. P. Uberuaga, and H. J \'o nsson. A climbing image nudged elastic band method for finding saddle points and minimum energy paths . J. Chem. Phys. , 113(22):9901--9904, 2000

    work page 2000

  21. [21]

    E. A. Carter, G. Ciccotti, J. T. Hynes, and R. Kapral. Constrained reaction coordinate dynamics for the simulation of rare events . Chem. Phys. Lett. , 156(5):472--477, 1989

    work page 1989

  22. [22]

    J. J. Mortensen, A. H. Larsen, M. Kuisma, A. V. Ivanov, and et al. GPAW: An open Python package for electronic structure calculations . J. Chem. Phys. , 160(9):092503, 2024

    work page 2024

  23. [23]

    J. Kim, T. D. Baczewski, A. Beaudet, and et al. QMCPACK: an open source ab initio quantum Monte Carlo package for the electronic structure of atoms, molecules and solids . J. Phys.: Condens. Matter , 30:195901, 2018. /https://qmcpack.org/

    work page 2018

  24. [24]

    Perdew, K

    P. Perdew, K. Burke, and M. Ernzerhof. Generalized Gradient Approximation Made Simple . Phys. Rev. Lett. , 77:3865, 1996

    work page 1996

  25. [25]

    Giannozzi and et al

    P. Giannozzi and et al. QUANTUM ESPRESSO: a modular and open-source software project for quantum simulations of materials . J.Phys.: Condens.Matter , 21:395502, 2009. /https://www.quantum-espresso.org/

    work page 2009

  26. [26]

    Effective core potentials, 2024

    ECP community . Effective core potentials, 2024. https://pseudopotentiallibrary.org/

    work page 2024

  27. [27]

    Hlo z n \' y

    A. Hlo z n \' y . FT-MLP: MoS _ 2 S defect- github repository , 2026. The MLP is available at: https://github.com/AdamHlo/mos2-vacancy-mlp

    work page 2026

  28. [28]

    J. Heyd, G. E. Scuseria, and M. Ernzerhof. Hybrid functionals based on a screened coulomb potential. J. Chem. Phys. , 118:8207–8215, 2003

    work page 2003

  29. [29]

    J. Heyd, G. E. Scuseria, and M. Ernzerhof. Erratum: Hybrid functionals based on a screened coulomb potential. J. Chem. Phys. , 124:219906, 2006

    work page 2006

  30. [30]

    Mace foundation models

    Batatia, I. Mace foundation models. https://github.com/ACEsuit/mace-foundations

    work page

  31. [31]

    N. Ch. Bennett, G. Wang, A. Annaberdiyev, A. C. Melton, L. Shulenburger, and L. Mitas. A new generation of effective core potentials from correlated calculations: 2nd row elements . J. Chem. Phys. , 149:104108, 2018

    work page 2018

  32. [32]

    G. Wang, B. Kincaid, Zhou. H., A. Annaberdiyev, N. Ch. Bennett, J. T. Krogel, and L. Mitas. A new generation of effective core potentials from correlated and spin-orbit calculations: Selected heavy elements . J. Chem. Phys. , 157:05410, 2022

    work page 2022

  33. [33]

    Beyond the locality approximation in the standard diffusion monte carlo method

    Casula, M. Beyond the locality approximation in the standard diffusion monte carlo method. Phys. Rev. B , 74:1--4, 2006

    work page 2006