Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Universal machine-learning interatomic potentials fail at melt-quench amorphous structure generation because their energy-volume and pressure errors cause unphysical expansion; revised NVT quench and pressure-targeted fine-tuning restore re

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 11:11 UTC pith:GFUBPF4Y

load-bearing objection Convincing failure diagnosis for uMLIP melt-quench, but the claimed general cure is under-supported—the short NPT stage likely inherits the starting crystal-like volume. the 3 major comments →

arxiv 2606.16385 v2 pith:GFUBPF4Y submitted 2026-06-15 cond-mat.mtrl-sci

Melt-Quench Failures and Practical Solutions for Universal Machine-Learning Interatomic Potentials in Amorphous Structure Generation

Shuwei Li , Yuqi An , Xingyu Guo , Wenqiang Yang , Zhenbin Wang This is my paper
classification cond-mat.mtrl-sci
keywords universal machine-learning interatomic potentialsmelt-quench molecular dynamicsamorphous IrO2energy-volume responsevirial pressureNVT-quench protocolpressure-targeted fine-tuningamorphous structure generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that the conventional NPT melt-quench protocol, when driven by eight leading universal machine-learning interatomic potentials, produces unphysically expanded amorphous structures, not because energies or forces are inaccurate but because the models misjudge how energy changes with volume and how virial pressure responds. Using amorphous IrO2 as a diagnostic, it shows all eight models predict densities of 1-4 g/cm3 against a DFT reference of 10.04 g/cm3. The paper then demonstrates two fixes: fine-tuning with a stress-weighted loss, and, more generally, a revised protocol that keeps the cell volume fixed during the high-temperature quench and only applies NPT at low-temperature equilibration. Across 30 chemically diverse materials, the revised protocol reduces the mean absolute density error from 2.46 to 0.35 g/cm3 against AIMD references. The significance is that validation of universal potentials must include energy-volume curves and pressure, not just energy and force benchmarks.

Core claim

The central claim is that accurate energies and forces are necessary but not sufficient for stable NPT melt-quench dynamics; what matters is the sign of the energy-volume error and the fidelity of the virial pressure. The paper shows that a model can be more accurate than a system-specific reference on every instantaneous quantity and still drive catastrophic expansion, while a less accurate model with the correct E-V sign stays stable. It also shows that even models whose energy-volume curves match DFT can still fail because they systematically overestimate pressure. The two remedies — pressure-targeted fine-tuning and an NVT-quench/NPT-equilibration protocol — both recover amorphous IrO2 d

What carries the argument

The load-bearing objects are the energy-volume (E-V) curve and the virial pressure computed from the stress tensor. The paper's central diagnostic is the sign of the E-V error under volume expansion, not its magnitude, and the central fix is a simulation protocol in which the high-temperature quench runs in the NVT ensemble (fixed cell volume) and only the low-temperature equilibration runs in NPT, so incorrect high-temperature E-V or pressure responses cannot drive volume evolution. A second mechanism is pressure-targeted fine-tuning, which reweights the loss toward the stress term to repair pressure predictions without losing energy/force accuracy.

Load-bearing premise

The broad claim of general applicability rests on treating the 30 reference densities — some from computer simulation, some from experiment, some from the corresponding crystals — as fair targets for the potential; if those references are not directly comparable with what the simulation is trying to reproduce, the reported improvement is partly an artifact of the reference list.

What would settle it

Rerun the conventional and revised melt-quench protocols with any of the eight potentials on a material with an accurately known experimental amorphous density outside the 30-material set (e.g., fused silica or amorphous selenium). If the conventional protocol produces a near-reference density, or the revised protocol misses it by more than the reported error scale, the generality claim is wrong. Alternatively, compute the DFT energy along one of the uMLIP's expansion paths: if DFT shows an energy minimum near the reference density where the potential predicts a monotonic decrease, the propose

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Benchmarks for universal machine-learning interatomic potentials should include energy-volume curves and pressure parities, not just energy and force errors, because the conventional metrics miss the failure mode.
  • Published amorphous structures generated with universal potentials via conventional NPT melt-quench should be re-examined; the expanded densities are an artifact of the protocol, not the material.
  • The revised NVT-quench/NPT-equilibration protocol is a drop-in fix that requires no additional training data and should become the default for uMLIP-driven amorphous structure generation.
  • When fine-tuning is feasible, stress-weighted fine-tuning can correct volume behavior for a specific material while preserving energy and force accuracy.
  • The failure generalizes across oxides, nitrides, sulfides, halides, and ternary oxides, so any uMLIP used for non-equilibrium or volume-varying simulations should first pass an E-V/pressure validation.

Where Pith is reading between the lines

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

  • If the sign-of-E-V-error mechanism is general, uMLIP training sets should deliberately include expanded, low-density amorphous configurations so the models learn the correct energetic penalty for volume growth.
  • The revised protocol fixes density by suppressing volume degrees of freedom at high temperature; properties that depend on genuine high-temperature volume response, such as thermal expansion or pressure-induced transformations, would still require pressure-accurate models.
  • The 30-material benchmark mixes AIMD, experimental, and crystalline reference densities; the qualitative failure is robust, but the quantitative 0.35 g/cm3 MAE is only defined on the 19 AIMD-referenced materials, and extending it to all 30 assumes those references are comparable.
  • A direct extension would be to test whether the same NVT-quench fix applies to other non-equilibrium NPT processes, such as rapid cooling of metallic glasses or nucleation from the melt, where the same pressure-error mechanism could distort the outcome.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper evaluates eight universal machine-learning interatomic potentials (GRACE, ORB, SevenNet, MACE, MatterSim, UMA, UPET, NequIP) for melt–quench generation of amorphous IrO2. It reports that all eight fail under the conventional NPT quench protocol, yielding densities of 1–4 g/cm3 versus an AIMD (PBE) reference of 10.04 g/cm3. Using DFT benchmarks on trajectory snapshots, the authors argue that the failure is caused by incorrect energy–volume responses and/or virial-pressure overestimation, rather than by poor pointwise energy/force accuracy. They propose two remedies: pressure-targeted fine-tuning of UPET, and a revised NVT-quench/NPT-equilibration protocol. They claim the revised protocol restores IrO2 densities for all eight models and reduces the density MAE from 2.46 to 0.35 g/cm3 on a 30-material benchmark (AIMD-referenced subset, GRACE runs).

Significance. If the central claims hold, this is a useful and timely contribution to a practical problem: universal MLIPs are increasingly used for amorphous structure generation, and standard energy/force benchmarks are shown to be insufficient for NPT melt–quench simulations. The diagnosis that the sign of the E–V error and the fidelity of virial pressure are controlling factors is more informative than aggregate energy/force MAEs alone. Strengths include repeated independent IrO2 runs, DFT benchmarking of trajectory snapshots, an eight-model comparison on IrO2, and an out-of-sample 30-material test of the revised protocol. However, the 30-material test uses only one uMLIP (GRACE), mixes reference types of unequal physical meaning, and is subject to a volume-heredity concern that requires additional convergence evidence before the protocol can be accepted as a general physical cure.

major comments (3)
  1. [§2.4 and Methods §4.1; Fig. 6c–e] The central claim that the revised NVT-quench/NPT-equilibration protocol recovers AIMD-consistent densities is not yet established because the initial cell volume is set externally and the final NPT stage is very short. For IrO2, the simulations start from the rutile crystal cell; for the 30-material set, initial packings are built at 1.05× the Materials Project crystal density. Since the quench is performed in NVT, the density entering the 10 ps NPT equilibration is predetermined to be near the crystal density. For GRACE, ORB, SevenNet, MACE, MatterSim, and UMA, the E–V curves in Fig. 3c–h are monotonically decreasing over the tested range, so no stable volume minimum exists at 300 K; a longer NPT run would be expected to continue expanding. The paper reports no density-versus-time curves for the final NPT stage, no longer-NPT tests, and no dependence on the starting cell volume. The re
  2. [Table S2 and Fig. 6e–f] The 30-material benchmark mixes reference types and uses only GRACE. The headline AIMD-referenced MAE includes the LDA-based WO3 entry, and 11 of the 30 entries use experimental or crystalline-phase reference densities. For amorphous structure generation, a crystalline reference density is not a commensurate ground truth: amorphous densities are generally lower than crystal densities, and the revised protocol deliberately starts from crystal-density packings, so a model that stays near the starting volume will appear to succeed against crystal references even if its amorphous density is wrong. In addition, all 30-material runs are performed with only one uMLIP (GRACE), so the abstract/conclusion statements that the failure and remedy are general to uMLIPs are broader than the evidence. Please report per-reference-source metrics, remove or clearly separate non-PBE-AIMD entries from the he
  3. [§2.3, Figs. 4–5] The pressure-targeted fine-tuning demonstration is an in-domain retraining check rather than an independent validation. UPET-FT is trained on AIMD IrO2 configurations and then evaluated on the same material, so it shows that a uMLIP can be corrected when abundant target-domain AIMD data are available, but it does not establish fine-tuning as a general remedy. If this is retained as one of the two central practical solutions, please either restrict the claim to a proof-of-concept or add a second uMLIP and a held-out material to demonstrate transferability.
minor comments (4)
  1. [Methods, Eq. (4)] The loss expression contains '10.0&&' where the stress-loss coefficient should be defined. Please replace with an explicit term such as 10.0 L_S.
  2. [Fig. 6c] The density values for the revised protocol are reported to two decimals and agree closely with AIMD, but no error bars are visible for the individual uMLIP bars in the figure as printed. Please clarify the statistics for these revised-protocol runs.
  3. [Methods §4.1] The pressure control uses the ASE Berendsen-style barostat. Since the paper’s central concern is NPT volume dynamics, please state whether the conclusions are expected to hold with a more rigorous barostat (e.g., Parrinello–Rahman) and whether any thermostat/barostat coupling parameters were varied.
  4. [Data availability] The MTP used as a domain-specific reference is cited as 'In preparation' (Ref. 25). For reproducibility, the MTP training data, active-learning details, and final potential should be made available in a public repository.

Circularity Check

0 steps flagged

No circularity found: the benchmark findings are empirical, the 30-material MAE is not fit to the references, and the same-group MTP/UPET-FT items are supported within the paper.

full rationale

The paper's derivation chain is not circular. The central failure claim is an empirical observation: all eight uMLIPs under NPT melt-quench yield 1-4 g/cm3 vs AIMD 10.04 g/cm3 (Fig. 1b), benchmarked against DFT/AIMD, not derived from any fitted parameter. The E-V diagnosis (Fig. 3) and pressure parities (Fig. 4) are direct comparisons to DFT/AIMD references; no equation identifies a uMLIP prediction with a reference value. The 30-material revised-protocol MAE (0.35 g/cm3 AIMD-referenced) is computed from GRACE simulations and literature references; the protocol was not fit to those densities, so the improvement is out-of-sample. The MTP is cited to an in-preparation same-group paper, but the current manuscript specifies its active-learning training and validates it against DFT (Fig. 2, S2, S3), so the self-citation is not load-bearing. UPET-FT is fine-tuned on AIMD IrO2 configurations and then re-tested on IrO2; this is an in-domain consistency demonstration, not an independent prediction, and no claim of cross-system prediction is made for it. The strongest validity concern is the revised protocol's NVT-fixed volume plus only 10 ps final NPT, so reported densities could partly inherit the crystal-based starting volume (Methods 4.1); this missing convergence/initial-volume control is an omitted proof for the 'physical cure' interpretation, but it is a confounding-control problem, not a definitional circularity, because the paper does not define the predicted density as equal to its input volume by any equation. Therefore no specific circular step can be exhibited.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim is empirical; it rests on the quality of DFT/AIMD references, a small set of hand-chosen protocol parameters (fine-tuning weights, packing densities), and the assumption that the tested uMLIP implementations and ASE barostat are representative. No new physical entities are postulated.

free parameters (3)
  • UPET fine-tuning loss weights (wE, wF, wS) = 0.1, 0.1, 10.0
    Hand-chosen to weight stress/pressure highly during fine-tuning; this is the core of the pressure-targeted remedy and was not systematically optimized.
  • MTP training loss weights (energy:force:stress) = 1 : 0.001 : 0.01
    Used to train the system-specific MTP that serves as the successful reference in the E-V/pressure paradox; a hyperparameter in the in-preparation Ref. 25.
  • Initial packing density multiplier for 30-material benchmark = 1.05 × crystal density
    All 30 benchmark initial configurations were generated with Packmol at 1.05 times the crystalline density; this starting volume is a hand-chosen protocol parameter for the NVT quench.
axioms (4)
  • domain assumption PBE-DFT/AIMD provides the correct reference for amorphous IrO2 density and energetics.
    The AIMD density of 10.04 g/cm3 is treated as ground truth in §2.1; PBE is known to overestimate volumes, and the optB86b reference is 10.70 g/cm3, so the reference itself is functional-dependent.
  • domain assumption The MTP of Ref. 25 faithfully reproduces AIMD behavior of amorphous IrO2.
    Used as the successful system-specific MLIP in the paradox argument in §2.2; its construction is in an in-preparation manuscript and cannot be independently verified here.
  • domain assumption Literature reference densities for the 30-material benchmark are commensurate with PBE-level uMLIP predictions.
    Table S2 mixes AIMD (mostly PBE, one LDA), experimental, and crystalline-phase densities; comparing these to GRACE NVT-quench outputs assumes they are valid proxies for the target amorphous density.
  • domain assumption The ASE Berendsen-style barostat/Langevin thermostat with a 1 fs timestep provides physically meaningful NPT/NVT dynamics for uMLIP-driven melt-quench.
    All uMLIP MD uses ASE's default barostat; the failure and fix are demonstrated within this integrator/barostat choice, which may not be equivalent to the Parrinello-Rahman dynamics used in AIMD.

pith-pipeline@v1.3.0-alltime-deepseek · 18316 in / 17953 out tokens · 189551 ms · 2026-08-02T11:11:21.681571+00:00 · methodology

0 comments
read the original abstract

Generating experimentally relevant amorphous structures via melt-quench molecular dynamics is prohibitively expensive at the first-principles level. Universal machine-learning interatomic potentials (uMLIPs) could accelerate such simulations, but their reliability under non-equilibrium conditions remains unclear. Here, we examine eight leading uMLIPs for generating amorphous IrO2, using this electrocatalytically relevant oxide as a diagnostic case. Under the conventional melt-quench protocol, all models yield unphysically expanded structures with densities of 1-4 g/cm3, far below the ab initio molecular dynamics (AIMD) reference value of 10.04 g/cm3. Comparisons against ab initio references show that accurate energies and forces alone do not ensure stable NPT dynamics; correct energy-volume responses and pressure predictions are also essential. We identify two practical remedies: pressure-targeted fine-tuning and a revised NVT-quench/NPT-equilibration protocol that avoids unphysical volume expansion without additional ab initio training data. Both recover IrO2 densities and local structures consistent with AIMD. Across 30 chemically diverse materials, the volume-expansion failure proves general, and the revised protocol substantially improves density predictions, reducing the AIMD-referenced MAE from 2.46 to 0.35 g/cm3. This work establishes practical validation criteria and simulation strategies for robust uMLIP-driven amorphous structure generation.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Predicting Novel Stable Materials for Experimental Synthesis

    cond-mat.mtrl-sci 2026-07 conditional novelty 4.0

    A hierarchical screening protocol using PBE phase diagrams, ML interatomic potentials, and SCAN refinement reduces 894 computationally stable materials to 25 high-confidence experimental synthesis targets.

Reference graph

Works this paper leans on

6 extracted references · 2 canonical work pages · cited by 1 Pith paper

  1. [2]

    Failure of uMLIPs under conventional melt–quench simulations The melt–quench method is widely used for generating amorphous structures[22,23]

    Results and Discussion 2.1. Failure of uMLIPs under conventional melt–quench simulations The melt–quench method is widely used for generating amorphous structures[22,23]. In this approach (Figure 1a), melting is performed in the NVT ensemble, followed by quenching and equilibration in the NPT ensemble to allow volume evolution during cooling. We calculate...

    arXiv 1980

  2. [4]

    Recent Progress of Amorphous Nanomaterials,

    Methods 4.1 Molecular Dynamics Simulations uMLIP-driven molecular dynamics: Melt–quench molecular dynamics (MD) simulations driven by universal machine-learning interatomic potentials (uMLIPs) were performed using the Atomic Simulation Environment (ASE)[26]. Temperature was controlled with a Langevin thermostat in the canonical (NVT) ensemble, whereas pre...

    work page doi:10.48550/arxiv.2510.22999 2023

  3. [5]

    Density Functional Modeling of Structural and Electronic Properties of Amorphous High Temperature Oxides,

    V. B. Sulimov, D. C. Kutov, A. V. Sulimov, F. V. Grigoriev, and A. V. Tikhonravov, “Density Functional Modeling of Structural and Electronic Properties of Amorphous High Temperature Oxides,” Journal of Non-Crystalline Solids 578 (2022): 121170, https://doi.org/10.1016/j.jnoncrysol.2021.121170. 18. A. F. Harper, K. Iwanowski, W. C. Witt, M. C. Payne, and M...

    arXiv 2022

  4. [17]

    Universal Machine Learning Interatomic Potentials Are Ready for Phonons,

    A. Loew, D. Sun, H.-C. Wang, S. Botti, and M. A. L. Marques, “Universal Machine Learning Interatomic Potentials Are Ready for Phonons,” npj Computational Materials 11, no. 1 (2025): 178, https://doi.org/10.1038/s41524-025-01650-1. 18. S. Zhu, D. Sarıtürk, and R. Arróyave, “Machine Learning Potentials for Alloys: A Detailed Workflow to Predict Phase Diagra...

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.06231 2025

  5. [32]

    Amorphous-to-Crystal Transition in Quasi-Two-Dimensional MoS2: Implications for 2D Electronic Devices,

    M. Krbal, V. Prokop, A. A. Kononov, et al., “Amorphous-to-Crystal Transition in Quasi-Two-Dimensional MoS2: Implications for 2D Electronic Devices,” ACS Applied Nano Materials 4, no. 9 (2021): 8834–8844, https://doi.org/10.1021/acsanm.1c01504. 33. Y. Saito, C. Kaito, and T. Naiki, “Structure of Thin Amorphous MoO3 Films Prepared by Vacuum-Deposition,” Jou...

    work page doi:10.1021/acsanm.1c01504 2021

  6. [33]

    Pushing the Limits of Unconstrained Machine-Learned Interatomic Potentials,

    F. Bigi, P. Pegolo, A. Mazitov, J. Schmidt, and M. Ceriotti, “Pushing the Limits of Unconstrained Machine-Learned Interatomic Potentials,” Pushing the limits of unconstrained machine-learned interatomic potentials, arXiv 2026, https://doi.org/10.48550/arXiv.2601.16195. 34. C. W. Tan, M. L. Descoteaux, M. Kotak, et al., “High-Performance Training and Infer...

    Pith/arXiv arXiv doi:10.48550/arxiv.2601.16195 2026