Amorphous materials as a frontier challenge for universal interatomic potentials

Natascia L. Fragapane; Volker L. Deringer

Current universal machine-learned potentials struggle with amorphous solids, a frontier that crystal-focused benchmarks miss.

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 · grok-4.5

2026-07-14 01:27 UTC pith:GTDNQ6BM

load-bearing objection Timely benchmarking pitch that amorphous solids expose transferability gaps in crystal-trained universal MLIPs—but from the abstract alone we cannot yet judge how representative the curated set is. the 3 major comments →

arxiv 2607.11384 v1 pith:GTDNQ6BM submitted 2026-07-13 cond-mat.mtrl-sci physics.comp-ph

Amorphous materials as a frontier challenge for universal interatomic potentials

Natascia L. Fragapane , Volker L. Deringer This is my paper

classification cond-mat.mtrl-sci physics.comp-ph

keywords machine-learned interatomic potentialsamorphous materialstransferabilityfoundational modelsbenchmarkingfine-tuningdisordered solids

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.

Pre-trained machine-learned interatomic potentials (MLIPs) have become standard tools for materials modelling, yet most early models and their benchmarks were built around ordered crystals. This paper argues that the amorphous state is a distinct, central challenge for the next generation of universal MLIPs. The authors evaluate mainstream pre-trained models on a curated reference set of canonical amorphous systems, checking both structures and properties, and find clear transferability limits that crystal-only tests do not reveal. They also examine fine-tuning strategies aimed at disordered phases. If the claim holds, researchers modelling glasses, amorphous semiconductors, and other disordered functional materials will need more carefully designed training data and validation protocols rather than assuming crystal-trained models will simply transfer.

Core claim

Systematic evaluation of current mainstream pre-trained MLIPs on a curated set of canonical amorphous systems shows that many models fail to transfer reliably to non-crystalline solids; crystal-focused benchmarks therefore do not capture the performance limits that matter for disordered materials.

What carries the argument

A benchmarking framework built on a curated reference dataset of canonical amorphous systems, together with structure and property validation metrics, used to expose transferability gaps and to test fine-tuning strategies for disordered phases.

Load-bearing premise

The chosen set of canonical amorphous systems and the structure/property metrics used for validation are representative enough of the broader space of disordered solids that the observed failures can be treated as a general frontier challenge rather than artifacts of a narrow test set.

What would settle it

A mainstream pre-trained MLIP that, without specialized fine-tuning, reproduces reference structures and key properties (for example radial distribution functions, densities, or mechanical/thermal observables) for the paper's canonical amorphous systems at the same accuracy level it achieves on crystalline benchmarks.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Timely benchmarking pitch that amorphous solids expose transferability gaps in crystal-trained universal MLIPs—but from the abstract alone we cannot yet judge how representative the curated set is.

read the letter

Punchline: this is a methods/benchmarking paper claiming that mainstream pre-trained MLIPs, built largely on ordered crystals, transfer poorly to amorphous solids, and that the field needs a dedicated amorphous reference set plus fine-tuning guidance. From the abstract that claim is plausible and useful; we cannot yet score the evidence.

What is actually new is the package: a curated “canonical” amorphous reference dataset, a structure/property validation framework, a systematic sweep of current foundational models, and a look at fine-tuning for disordered phases. Crystal-centric benchmarks have been the default. Calling out the amorphous gap and shipping a reference set is the right kind of contribution if the data and protocols hold up. The design is structurally clean—external evaluation of pre-trained models on held-out amorphous references—so circularity is low by construction. The problem statement is also clean: universal MLIPs are sold as general-purpose; checking them on disordered systems is an obvious next step many groups have only done ad hoc.

Soft spots, in proportion: we only have the abstract. The stress-test concern lands at this stage. The jump from “models fail on our curated set” to “the amorphous state is a frontier challenge for universal MLIPs” depends on how representative the systems, chemistries, preparation routes, and metrics are. If the set is narrow, or if metrics punish missing long-range order by construction, the headline overreaches. Soundness, baselines, error bars, model list, and data release are uncheckable without the full text. That is a limitation of this read, not a proven flaw in the paper.

Who it is for: people building or using foundational MLIPs, and groups working on amorphous functional materials. It deserves a serious referee rather than a desk reject—the question is important enough and the design is non-circular. I would not bring it to a general reading group until the dataset and tables are inspectable; for an MLIP-validation group, yes. Send to peer review; insist on full inventory of systems, quantitative crystal-versus-amorphous comparisons, and open data.

Referee Report

3 major / 3 minor

Summary. The manuscript argues that the amorphous state is a central challenge for universal (pre-trained/foundational) machine-learned interatomic potentials (MLIPs). Early models and benchmarks have focused on ordered crystals, leaving transferability to non-crystalline solids unclear. The authors report a systematic evaluation of mainstream pre-trained models on a curated reference dataset of canonical amorphous systems, with structure and property validation; they identify transferability limitations, examine fine-tuning strategies for disordered phases, and introduce a benchmarking framework intended to guide next-generation training sets and models for amorphous functional materials.

Significance. If the evaluation is rigorous and the reference set is representative, the work would fill a genuine gap: crystal-centric MLIP benchmarks can miss failures on disordered solids, and a reusable amorphous benchmarking framework would be of practical value to the community. Explicit investigation of fine-tuning for disordered phases is also useful. The abstract-level contribution is therefore potentially significant for materials modelling of glasses and amorphous functional materials; significance cannot be confirmed without the full methods, data inventory, and quantitative results.

major comments (3)

Abstract (central claim): The assertion that “the amorphous state is indeed a central challenge for future universal MLIPs” is load-bearing and rests on generalization from a curated “canonical” reference set. The abstract supplies no inventory of systems, chemistries, preparation protocols (e.g., melt-quench vs. other routes), or property metrics. Without that inventory and a clear argument for representativeness of disordered solids more broadly, the step from “models fail on our set” to “amorphous materials are a frontier challenge for the field” cannot be assessed and may over-generalize from a narrow test suite.
Abstract (evaluation design): Transferability limitations are claimed relative to crystal-focused practice, yet the abstract does not state whether the same models were evaluated side-by-side on crystalline counterparts of the same chemistries, nor whether failures track absence of long-range order versus chemistries already outside the pre-training distribution. A crystal-versus-amorphous comparison (or an explicit statement that crystalline performance is already known and adequate) is needed for the “amorphous-specific challenge” framing to hold.
Abstract (results support): The abstract asserts “systematic evaluation,” “identifies limitations,” and a “benchmarking framework” without any quantitative error tables, model list, baselines, error bars, or data-exclusion criteria. Those elements are load-bearing for the central claim; the full manuscript must present them in a form that allows independent judgment of effect sizes and of whether reported failures are robust rather than metric- or protocol-specific artifacts.

minor comments (3)

Abstract wording: “canonical amorphous systems” and “validation for structures and properties” are left undefined; even a brief parenthetical list of system classes (e.g., oxides, metals, semiconductors) and metric families would improve clarity for readers scanning the abstract.
Abstract: “mainstream models” should be named or at least scoped (architecture families / public foundational MLIPs) so the claimed coverage is checkable.
Abstract: Fine-tuning strategies are mentioned as investigated; a one-phrase indication of what was varied (data composition, loss terms, freeze/unfreeze policy) would help readers judge novelty relative to existing fine-tuning practice.

Circularity Check

0 steps flagged

Abstract-only benchmarking study of external pre-trained MLIPs on amorphous systems shows no circular derivation; evaluation is independent of model training by construction.

full rationale

Only the abstract is available. It describes a systematic evaluation of existing mainstream pre-trained MLIPs on a curated reference dataset of canonical amorphous systems, plus structure/property validation and fine-tuning investigations. No equations, fitted parameters, uniqueness theorems, or load-bearing self-citations appear in the provided text. The central claim—that amorphous solids pose a transferability challenge for universal MLIPs—is presented as an empirical finding from testing external models against independent reference data, not as a quantity derived from or fitted to the same inputs used as targets. Residual ordinary self-citation risk for prior amorphous datasets cannot be assessed without the full text and does not constitute circularity under the stated rules. Score 0 is therefore the correct honest finding for an abstract-only review of a benchmarking paper.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

Abstract-only review: free parameters and invented entities cannot be enumerated from equations or fits. The work rests on standard domain assumptions of atomistic MLIP modelling and on the representativeness of a curated amorphous reference set. No new physical entities are postulated; the contribution is methodological (benchmark + evaluation + fine-tuning).

axioms (3)

domain assumption Pre-trained foundational MLIPs trained primarily on crystalline data are expected to be transferable to other phases unless shown otherwise.
Implicit baseline the abstract tests against; standard in the universal-MLIP literature.
ad hoc to paper A curated set of “canonical” amorphous systems plus structure/property validation is a sufficient probe of transferability to non-crystalline solids.
The paper’s central generalization from its benchmark to “the amorphous state as a frontier challenge” depends on this representativeness claim.
domain assumption Standard interatomic-potential evaluation metrics (structural and property agreement with reference amorphous data) are meaningful for model quality.
Ordinary computational materials science practice; not derived in the abstract.

pith-pipeline@v1.1.0-grok45 · 6061 in / 2298 out tokens · 22103 ms · 2026-07-14T01:27:17.949773+00:00 · methodology

0 comments

read the original abstract

Pre-trained or 'foundational' machine-learned interatomic potentials (MLIPs) are now widely used in materials modelling. However, early pre-trained models and benchmarks have largely focused on ordered, crystalline structures, and their transferability to non-crystalline solids remains unclear. Here, we show that the amorphous state is indeed a central challenge for future universal MLIPs, based on a systematic evaluation of current mainstream models in this domain. We introduce a benchmarking framework built on a curated reference dataset of canonical amorphous systems, as well as validation for structures and properties. Our study identifies limitations in the transferability of many current pre-trained models and investigates fine-tuning strategies tailored to disordered phases. Together, our results can facilitate future applications of MLIPs in the fast-growing field of amorphous functional materials, and they provide guidance for designing next-generation training datasets and transferable atomistic models.

Amorphous materials as a frontier challenge for universal interatomic potentials

Core claim

What carries the argument

Load-bearing premise

What would settle it

discussion (0)