Pith. sign in

REVIEW 6 major objections 6 minor 1 cited by

Crystal-generating models that scale well fail predictably: their out-of-distribution error follows a cube-root scaling law that in-distribution fits can forecast.

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

2026-08-03 02:59 UTC pith:4ZWLPNZ4

load-bearing objection Solid, reproducible benchmark with useful diagnostics, but the 'forecastable frontier' claim rests on four narrow OOD radii and is likely a scale-calibration artifact. the 6 major comments →

arxiv 2602.09309 v2 pith:4ZWLPNZ4 submitted 2026-02-10 cond-mat.mtrl-sci cond-mat.mes-hallcs.LGphysics.atm-clus

How Far Can You Grow? Characterizing the Extrapolation Frontier of Graph Generative Models for Materials Science

Can Polat , Erchin Serpedin , Mustafa Kurban , Hasan Kurban This is my paper
classification cond-mat.mtrl-sci cond-mat.mes-hallcs.LGphysics.atm-clus
keywords extrapolation frontiergenerative modelsmaterials sciencenanoparticlescrystal generationscaling lawout-of-distribution generalizationgraph neural networks
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 generative models for crystalline materials have a measurable 'extrapolation frontier'—a size beyond which generated structures become unreliable—and that this frontier is not chaotic but follows a power law. To test this, it builds RADII, a benchmark of about 75,000 sphere-cut nanoparticles from ten materials across 25 radii, with splits that keep test sizes and orientations strictly out of the training set. Five generative models are evaluated on a geometry-only task: given a unit cell, a target radius, and the exact atom count and species ordering, produce coordinates. The central finding is that well-behaved models degrade by roughly 13% in global positional error beyond training radii, while local bond fidelity ranges from nearly unchanged to more than double; the pattern of failure is different for every architecture; and error grows with the cube root of atom count, so in-distribution fits predict out-of-distribution error. If correct, this gives researchers a way to forecast when a model will start producing unreliable structures, and it elevates output scale to a standard evaluation axis.

Core claim

On the paper's own terms, the discovery is that the extrapolation frontier is a diagnosable and forecastable quantity. When a generative model is conditioned on a unit cell, a target radius, and the atom count, its per-radius positional error follows a power law RMSD ~ N^alpha on in-distribution radii; for the architectures that generate coherent structures, alpha is consistently near 1/3 with near-perfect fit, meaning error grows with the nanoparticle's linear dimension. Fitting this law on training radii predicts error at larger and smaller out-of-distribution radii with near-zero residual, so performance at unseen sizes can be estimated without running the model there. The same data show

What carries the argument

The central object is the power-law scaling identity RMSD ~ N^alpha with alpha close to 1/3, fitted on in-distribution radii and used as a predictor for out-of-distribution error. The mechanism that makes the measurement possible is RADII's construction: spherical truncation of a fixed periodic lattice turns radius into a continuous scaling knob, while leakage-free splits (interleaved in-distribution radii, strictly exterior out-of-distribution radii, and quaternion-sampled orientations excluded against training) let per-radius error profiles be attributed to scale rather than to orientation or composition. The correspondence between predicted and ground-truth atoms is guaranteed by conditio

Load-bearing premise

The load-bearing premise is that the task is a meaningful generative extrapolation problem even though the ground truth is a deterministic spherical truncation of the input unit cell; the paper itself notes that a rule-based baseline given the lattice parameters would trivially achieve zero error, so all measured degradation belongs to how well fixed-sequence coordinate regressors approximate this deterministic map under a scale shift.

What would settle it

Train any architecture on RADII's training radii (8-28 Å) and evaluate at the held-out out-of-distribution radii (6, 7, 29, 30 Å). The claim predicts that if the in-distribution fit gives alpha ~ 1/3 with R^2 near 1, the out-of-distribution residual will be near zero (the paper reports 0.0008-0.0036 for well-behaved models); a model with alpha ~ 1/3 but residual comparable to DiffCSP's 0.118 would falsify the forecastability claim. Alternatively, finding a well-behaved model whose per-radius error grows faster than N^(1/3) on in-distribution radii and yet extrapolates well would break the prop

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

If this is right

  • If the cube-root law holds, a model's error at an unseen nanoparticle size can be estimated from a fit on training radii alone, turning frontier prediction into a cheap diagnostic.
  • Output scale should be reported as a standard evaluation axis for geometric generative models; a model that looks good at one fixed size may degrade at larger or smaller sizes.
  • Because no two architectures share the same failure sequence, multi-metric evaluation (global position, local bonds, radius of gyration, coordination) is necessary to know what breaks first.
  • The uniform surface/interior growth implies that extrapolation failures are bulk-scale geometric errors, not edge effects, so fixing surface handling alone will not close the frontier.
  • The finding that unstable models show poor fidelity across scales suggests that a single scaling-law check can serve as a sanity filter for new architectures.

Where Pith is reading between the lines

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

  • The alpha ~ 1/3 exponent implies that per-atom errors are not independent: if they were, RMSD would be flat in atom count rather than growing as N^(1/3). A direct measurement of error autocorrelation or a radial-contraction test would show whether the growth comes from a coherent spatial mode, such as slight shrinkage, rather than random noise.
  • Because the benchmark uses ideal sphere-truncated lattices without relaxation, a natural extension is to ask whether the same exponent survives for DFT-relaxed or Wulff-shaped references; if it does, the forecastability result would carry over to practical nanostructure prediction.
  • The conditioning removes composition and ordering from the task, so the frontier measured here is a lower bound on difficulty: a fully unconditional model must also extrapolate chemistry, and its frontier could appear at smaller sizes.
  • The forecastability claim suggests an inverse use: pick the largest training radius that keeps projected out-of-distribution error under a target threshold, turning the scaling law into a data-budget planner for future generative models.

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

6 major / 6 minor

Summary. The paper introduces RADII, a benchmark of roughly 75,000 spherical nanoparticles derived by deterministic truncation of ten crystalline unit cells at radii from 6 to 30 Å, with leakage-free orientation splits. Five generative models (CDVAE, DiffCSP, FlowMM, MatterGen, ADiT) are conditioned on the unit cell, radius, atom count, species sequence, and atom ordering, and evaluated on RMSD, bond-length MAE, surface/interior error, coordination correlation, and orientation stability. The main reported findings are: (i) a ~13% increase in global positional error for OOD radii, with larger variation in local bond fidelity; (ii) architecture-specific failure sequences; (iii) a power-law RMSD ~ N^α with α ≈ 1/3 for 'well-behaved' models, whose in-distribution fit is claimed to accurately predict out-of-distribution error, making extrapolation frontiers 'forecastable.'

Significance. If the central forecastability claim were established, the paper would provide a useful diagnostic axis for geometric generative models and a reproducible benchmark with practical value for nanomaterial design. Strengths include the deterministic, reproducible benchmark construction, the careful orientation-split protocol with explicit angular exclusion, the per-radius and surface/interior diagnostics, and the release of code and data. However, the forecastability claim is not yet supported: the OOD test consists of only four radii 1–2 Å outside the training range, the 'well-behaved' subset is selected post hoc using the same power-law fit that defines the law, and the observed α≈1/3 is consistent with a trivial global-scale artifact. Moreover, the task as formulated is deterministic reconstruction from fully specified inputs, so the significance for generative-model extrapolation in the usual sense is currently limited.

major comments (6)
  1. [Abstract; §4.1; §6] The abstract and conclusion state that 'all models degrade by ~13%' in global positional error. This is contradicted by §4.1, where DiffCSP's normalized OOD RMSD is 0.87 and MatterGen's raw RMSD is ~5905 Å (DiffCSP raw ID RMSD >3386 Å). Only ADiT (1.13), CDVAE (1.12), and FlowMM (1.13) support the 13% figure. The headline claim should be restricted to well-behaved models or corrected.
  2. [§3.1, Eqs. (1)–(3)] The task conditions on the unit cell (lattice vectors and basis positions), target radius, atom count, species sequence, and ordering. As the text admits, a rule-based baseline would 'trivially achieve zero error' because the ground truth is the deterministic spherical truncation of the input cell. The measured errors are therefore approximation errors of a deterministic coordinate-regression map under scale shift, not failures of generative modeling in the usual sense (no composition or ordering prediction, no sampling over structures). This limits the external validity of the 'generative model extrapolation frontier' framing and should be either reframed or supplemented with an unconditioned/assignment-free track.
  3. [§4.6, Fig. 4] The claim that in-distribution fits 'accurately predict' out-of-distribution error rests on only four OOD radii (6, 7, 29, 30 Å), which lie just 1–2 Å outside the training boundaries (8–28 Å). The reported residuals (ADiT 0.0008, FlowMM 0.0036) are evaluations of the same fitted line at nearby sizes, not independent forecasts. Extrapolation to substantially larger radii (e.g., 35–50 Å) or to a held-out size band, with uncertainty quantification, is needed to support the 'making frontiers forecastable' claim.
  4. [§4.6] The observed α≈1/3 is exactly what one expects if RMSD is proportional to nanoparticle radius R, since N^{1/3} ∝ R for constant-density spheres. This is consistent with a trivial global-scale artifact (e.g., a constant relative error) rather than a distinctive scaling law of generative models. Please report RMSD normalized by R (or N^{1/3}) versus R, and compare against a null model of global isotropic scaling, to distinguish the two interpretations.
  5. [§4.6] The taxonomy 'well-behaved' versus 'divergent' is introduced after fitting the same power laws whose exponents and residuals define the taxonomy. Selecting models based on α≈1/3 and small OOD residuals, and then claiming those models obey the law, is circular. Define the criterion a priori (e.g., raw ID RMSD below a threshold, R² threshold) or report results for the full model set without selection.
  6. [Abstract (last sentence)] The abstract claims that 'Scaling MatterGen to its published parameter count stabilizes sampling but does not close the frontier, while DiffCSP remains unstable at published scale.' No corresponding experiment, table, or section in the body reports such scaled runs. Either add the supporting results or remove the claim.
minor comments (6)
  1. [§3.3, Eq. (4)] The notation for Rtrain is ambiguous: '22,...,28' should be written explicitly (24, 26, 28) to make the count of 15 training radii clear.
  2. [§3.5.1, Eq. (10)] BondMAE compares globally sorted distance vectors, which may conflate distinct local environments. This is acknowledged in §5, but it should also be noted where the metric is introduced.
  3. [§3.5.3, §5] The frontier radius r★(m, τ) is defined but never instantiated. The statement in §5 that tabulating it is future work is honest, but the paper would be stronger if at least one illustrative table were included.
  4. [Fig. 3(c)] The caption reports 'OOD S/I gap: avg: 1.00×, min: 1.00×, max: 1.00×' with no decimals. Report actual numeric values with error bars, as the text claims changes bounded by ±0.003.
  5. [§4.1] The list of verification diagnostics for DiffCSP/MatterGen is useful, but 'consistent Å units' is vague; specify the exact unit checks performed (e.g., comparing against known bond lengths).
  6. [References] Several crystal-structure references (e.g., [60], [61], [63], [65], [69]) are generic handbook entries without page or table numbers; please provide specific data identifiers.

Circularity Check

1 steps flagged

Scaling-law forecastability is partly definitional: the 'well-behaved' class is selected by the OOD residuals the law is said to predict.

specific steps
  1. self definitional [Section 4.6 (Scaling Laws for Nanostructure Generation); cf. Abstract claim (iii)]
    "OOD residuals further distinguish predictable from unstable scaling: ADiT (0.0008) and FlowMM (0.0036) maintain near-zero residuals, meaning ID scaling accurately predicts OOD degradation, whereas DiffCSP (0.1184) and MatterGen (0.0503) diverge substantially."

    The paper first labels models as 'predictable' versus 'unstable' using their OOD residuals, then asserts that for models in the α≈1/3 regime 'performance at unseen sizes can therefore be estimated from ID fits alone.' Since admission to the well-behaved/predictable class is conditioned on near-zero OOD residuals, the claimed forecastability of that class is guaranteed by the selection rule rather than derived from the in-distribution fit. The ID power-law fit alone would not identify which models are forecastable; OOD data are used to draw the class boundary, and that post-hoc boundary is then presented as the law's predictive success.

full rationale

The raw extrapolation check for ADiT, CDVAE, and FlowMM is not itself circular: a power law is fit to in-distribution radii and evaluated at held-out OOD radii, and the near-zero residuals are genuine empirical evidence for those three models at the nearby OOD radii (6, 7, 29, 30 Å). The circularity enters at the generalization step. The paper's headline claim—'well-behaved models obey ... whose in-distribution fit predicts out-of-distribution error, making frontiers forecastable'—defines the well-behaved class using the OOD residuals that are supposed to be predicted. Section 4.6 identifies 'predictable from unstable scaling' by near-zero OOD residuals and then concludes that ID fits alone estimate OOD performance for α≈1/3 models; that conclusion is true by construction for the selected class. A separate benchmark-validity concern, flagged by the paper itself, is that the task is deterministic reconstruction: Section 3.1 states the ground-truth nanoparticle 'is the deterministic spherical truncation of the unit cell (Eq. 3)' and that a rule-based baseline 'would trivially achieve zero error,' and models receive the unit cell, radius, atom count, and full ordering. This makes the measured degradation a property of fixed-sequence coordinate regression under a scale shift, not generative extrapolation in the usual sense; I treat this as a weakness rather than a formal circularity step. The self-citations (C2NP, CrysMTM, earlier TiO2 work) are related-work or dataset references and are not load-bearing for the scaling-law argument; no uniqueness theorem is imported from the authors' prior work. Overall, the central 'forecastable frontier' claim is partially circular (selection by outcome), while the underlying per-model OOD measurements retain independent empirical content for a narrow extrapolation window.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claims rest on one fitted scaling exponent per model plus several task-design choices. The most consequential is the deterministic conditioning, which makes the 'generation' task a reconstruction test. The scaling law's forecastability is an extrapolation of a fitted line, not an independent prediction.

free parameters (2)
  • power-law exponent alpha per model = ADiT 0.334; CDVAE 0.335; FlowMM 0.342; DiffCSP 0.142; MatterGen -0.126
    Fitted to log RMSD vs log atom count on ID radii (§4.6, Figure 4); used as the basis for the claim that ID fit predicts OOD error.
  • power-law intercept per model = not tabulated in the paper
    The constant in RMSD ≈ C N^alpha is also fitted; without it the OOD 'prediction' is not defined.
axioms (4)
  • domain assumption The target nanoparticle is the unrelaxed spherical truncation of the bulk lattice (Eq. 3).
    Benchmark ground truth assumes no surface reconstruction, passivation, or relaxation; paper acknowledges in §5 that results may change under relaxation.
  • ad hoc to paper Conditioning on atom count, species sequence, and ordering preserves generative-model validity while guaranteeing one-to-one correspondence.
    This conditioning (§3.1, §3.5) turns the task into deterministic coordinate regression from the unit cell; the paper itself notes a rule-based baseline would achieve zero error, but still interprets results as characterizing generative extrapolation.
  • domain assumption Kabsch-aligned RMSD with known correspondence is the appropriate primary error measure.
    The paper sets aside assignment-free metrics (§3.5) and acknowledges BondMAE's sorted-distance limitation; if RMSD is dominated by global scale mismatch, the alpha≈1/3 law is a geometric artifact.
  • ad hoc to paper The 'well-behaved' vs 'divergent' split of models is meaningful and not post-hoc.
    Models with large raw ID errors (DiffCSP, MatterGen) are excluded from the alpha≈1/3 generalization claim; the same error metrics feed both the classification and the scaling-law conclusion.

pith-pipeline@v1.3.0-alltime-deepseek · 16059 in / 17273 out tokens · 161955 ms · 2026-08-03T02:59:24.930842+00:00 · methodology

0 comments
read the original abstract

Every generative model for crystalline materials harbors a critical structure size beyond which its outputs become unreliable; we call this the extrapolation frontier. Despite its consequences for nanomaterial design, this frontier has never been systematically measured. We introduce RADII, a radius-resolved benchmark of ~75,000 crystal-derived nanoparticle structures (33-11,298 atoms) that treats radius as a continuous scaling knob, tracing generation quality from in- to out-of-distribution under leakage-free splits. Each model is conditioned on target composition and atom count, isolating geometric extrapolation as the evaluation variable. RADII provides frontier-specific diagnostics: per-radius error profiles pinpoint each architecture's scaling ceiling, surface-interior decomposition separates boundary from bulk failures, and cross-metric sequencing reveals which aspect of structural fidelity breaks first. Benchmarking five state-of-the-art architectures, we find that: (i) well-behaved models degrade by ~13% in global positional error beyond training radii, while divergent models show poor fidelity across scales, with local bond fidelity ranging from negligible degradation to over 2x error growth; (ii) no two architectures share a failure sequence, revealing the frontier as a multi-dimensional surface shaped by model family; and (iii) well-behaved models follow the expected geometric scaling exponent alpha ~ 1/3, whose in-distribution fit predicts out-of-distribution error, making frontiers forecastable. Scaling MatterGen to its published parameter count stabilizes sampling but does not close the frontier, while DiffCSP remains unstable at published scale. These findings establish output scale as a first-class evaluation axis for geometric generative models. Code and data: https://github.com/KurbanIntelligenceLab/RADII.

Figures

Figures reproduced from arXiv: 2602.09309 by Can Polat, Erchin Serpedin, Hasan Kurban, Mustafa Kurban.

Figure 1
Figure 1. Figure 1: Detailed overview of RADII. (left) Unit cells of the materials—Ag, Au, CH [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: From primitive cell to radius-controlled nanoclusters. For each material in dataset, the panels show—left to right—the primitive unit cell followed by its canonical R = 6 Å and R = 30 Å nanoparticles. Materials are arranged from top to bottom in ascending order of the atom count in their R30 cluster, illustrating how coordination environments and bulk-like cores emerge with increasing radius. All views sha… view at source ↗
Figure 3
Figure 3. Figure 3: Extrapolation frontier across multiple dimensions of structural fidelity. (a) Global RMSD increases beyond the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Power-law scaling relationships quantify the extrap [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

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. VASP Plugins: Linking the Vienna ab-initio Simulation Package with Python

    cond-mat.mtrl-sci 2026-07 accept novelty 5.5

    A C++/pybind11 shared-memory plugin layer exposes VASP SCF and ionic data as NumPy arrays so Python can modify structure, forces, local potential, and occupancies in place.

Reference graph

Works this paper leans on

73 extracted references · 1 canonical work pages · cited by 1 Pith paper

  1. [1]

    Designing disorder into crystalline materials.Nature Reviews Chemistry, 4(12):657–673, 2020

    Arkadiy Simonov and Andrew L Goodwin. Designing disorder into crystalline materials.Nature Reviews Chemistry, 4(12):657–673, 2020

    2020

  2. [2]

    Tio2 nanostructures with different crystal phases for sensitive acetone gas sensors

    Shuang Cao, Ning Sui, Peng Zhang, Tingting Zhou, Jinchun Tu, and Tong Zhang. Tio2 nanostructures with different crystal phases for sensitive acetone gas sensors. Journal of Colloid and Interface Science, 607:357–366, 2022

    2022

  3. [3]

    Theory and simulation of crystal growth

    Andrea C Levi and Miroslav Kotrla. Theory and simulation of crystal growth. Journal of Physics: Condensed Matter, 9(2):299, 1997

    1997

  4. [4]

    Crystal growth and materials research in photovoltaics: progress and challenges.Journal of Crystal growth, 275(1-2):292–304, 2005

    Thomas Surek. Crystal growth and materials research in photovoltaics: progress and challenges.Journal of Crystal growth, 275(1-2):292–304, 2005

    2005

  5. [5]

    Density functional theory

    Maylis Orio, Dimitrios A Pantazis, and Frank Neese. Density functional theory. Photosynthesis Research, 102:443–453, 2009

    2009

  6. [6]

    Density functional tight binding.Philosoph- ical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 372(2011):20120483, 2014

    Marcus Elstner and Gotthard Seifert. Density functional tight binding.Philosoph- ical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 372(2011):20120483, 2014

    2011

  7. [7]

    Insights into current limitations of density functional theory.Science, 321(5890):792–794, 2008

    Aron J Cohen, Paula Mori-Sánchez, and Weitao Yang. Insights into current limitations of density functional theory.Science, 321(5890):792–794, 2008. 8

    2008

  8. [8]

    An efficient way to model complex magnetite: Assessment of scc-dftb against dft.The Journal of chemical physics, 150(9), 2019

    Hongsheng Liu, Gotthard Seifert, and Cristiana Di Valentin. An efficient way to model complex magnetite: Assessment of scc-dftb against dft.The Journal of chemical physics, 150(9), 2019

    2019

  9. [9]

    Comparison of reaxff, dftb, and dft for phenolic pyrolysis

    Tingting Qi, Charles W Bauschlicher Jr, John W Lawson, Tapan G Desai, and Evan J Reed. Comparison of reaxff, dftb, and dft for phenolic pyrolysis. 1. molec- ular dynamics simulations.The Journal of Physical Chemistry A, 117(44):11115– 11125, 2013

    2013

  10. [10]

    Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

    2021

  11. [11]

    Schnet–a deep learning architecture for molecules and materials

    Kristof T Schütt, Huziel E Sauceda, P-J Kindermans, Alexandre Tkatchenko, and K-R Müller. Schnet–a deep learning architecture for molecules and materials. The Journal of Chemical Physics, 148(24), 2018

    2018

  12. [12]

    Machine learning of accurate energy- conserving molecular force fields.Science Advances, 3(5):e1603015, 2017

    Stefan Chmiela, Alexandre Tkatchenko, Huziel E Sauceda, Igor Poltavsky, Kristof T Schütt, and Klaus-Robert Müller. Machine learning of accurate energy- conserving molecular force fields.Science Advances, 3(5):e1603015, 2017

    2017

  13. [13]

    Open catalyst 2020 (oc20) dataset and community challenges.Acs Catalysis, 11(10):6059–6072, 2021

    Lowik Chanussot, Abhishek Das, Siddharth Goyal, Thibaut Lavril, Muhammed Shuaibi, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, et al. Open catalyst 2020 (oc20) dataset and community challenges.Acs Catalysis, 11(10):6059–6072, 2021

    2020

  14. [14]

    E (n) equivariant graph neural networks

    Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. InInternational Conference on Machine Learning, pages 9323–9332. PMLR, 2021

    2021

  15. [15]

    Se (3)- transformers: 3d roto-translation equivariant attention networks.Advances in Neural Information Processing Systems, 33:1970–1981, 2020

    Fabian Fuchs, Daniel Worrall, Volker Fischer, and Max Welling. Se (3)- transformers: 3d roto-translation equivariant attention networks.Advances in Neural Information Processing Systems, 33:1970–1981, 2020

    1970

  16. [16]

    Spherenet: Learning spherical representations for detection and classification in omnidi- rectional images

    Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidi- rectional images. InProceedings of the European Conference on Computer Vision (ECCV), pages 518–533, 2018

    2018

  17. [17]

    Multimodal neural network- based predictive modeling of nanoparticle properties from pure compounds

    Can Polat, Mustafa Kurban, and Hasan Kurban. Multimodal neural network- based predictive modeling of nanoparticle properties from pure compounds. Machine Learning: Science and Technology, 5(4):045062, 2024

    2024

  18. [18]

    Molprop: Molecular property prediction with multimodal language and graph fusion.Journal of Cheminformatics, 16(1):56, 2024

    Zachary A Rollins, Alan C Cheng, and Essam Metwally. Molprop: Molecular property prediction with multimodal language and graph fusion.Journal of Cheminformatics, 16(1):56, 2024

    2024

  19. [19]

    Graphdf: A discrete flow model for molecular graph generation

    Youzhi Luo, Keqiang Yan, and Shuiwang Ji. Graphdf: A discrete flow model for molecular graph generation. InInternational Conference on Machine Learning, pages 7192–7203. PMLR, 2021

    2021

  20. [20]

    Space group con- strained crystal generation.arXiv preprint arXiv:2402.03992, 2024

    Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. Space group con- strained crystal generation.arXiv preprint arXiv:2402.03992, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Symmcd: Symmetry-preserving crystal generation with diffusion models.arXiv preprint arXiv:2502.03638, 2025

    Daniel Levy, Siba Smarak Panigrahi, Sékou-Oumar Kaba, Qiang Zhu, Kin Long Kelvin Lee, Mikhail Galkin, Santiago Miret, and Siamak Ravanbakhsh. Symmcd: Symmetry-preserving crystal generation with diffusion models.arXiv preprint arXiv:2502.03638, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Wyckoffdiff–a generative diffusion model for crystal symmetry.arXiv preprint arXiv:2502.06485, 2025

    Filip Ekström Kelvinius, Oskar B Andersson, Abhijith S Parackal, Dong Qian, Rickard Armiento, and Fredrik Lindsten. Wyckoffdiff–a generative diffusion model for crystal symmetry.arXiv preprint arXiv:2502.06485, 2025

    arXiv 2025

  23. [23]

    John Wiley & Sons, 2018

    Charles Kittel and Paul McEuen.Introduction to solid state physics. John Wiley & Sons, 2018

    2018

  24. [24]

    Gibbs–curie–wulff theorem in organic materials: a case study on the relationship between surface energy and crystal growth.Advanced Materials, 28(8):1697–1702, 2016

    Rongjin Li, Xiaotao Zhang, Huanli Dong, Qikai Li, Zhigang Shuai, and Wen- ping Hu. Gibbs–curie–wulff theorem in organic materials: a case study on the relationship between surface energy and crystal growth.Advanced Materials, 28(8):1697–1702, 2016

    2016

  25. [25]

    Nanoparticle shapes by using wulff constructions and first-principles calculations.Beilstein Journal of Nanotechnology, 6(1):361–368, 2015

    Georgios D Barmparis, Zbigniew Lodziana, Nuria Lopez, and Ioannis N Reme- diakis. Nanoparticle shapes by using wulff constructions and first-principles calculations.Beilstein Journal of Nanotechnology, 6(1):361–368, 2015

    2015

  26. [26]

    Kinetic and thermo- dynamic modified wulff constructions for twinned nanoparticles.The Journal of Physical Chemistry C, 117(31):15859–15870, 2013

    Emilie Ringe, Richard P Van Duyne, and Laurence D Marks. Kinetic and thermo- dynamic modified wulff constructions for twinned nanoparticles.The Journal of Physical Chemistry C, 117(31):15859–15870, 2013

    2013

  27. [27]

    Quantum dots and their multimodal applications: a review.Materials, 3(4):2260–2345, 2010

    Debasis Bera, Lei Qian, Teng-Kuan Tseng, and Paul H Holloway. Quantum dots and their multimodal applications: a review.Materials, 3(4):2260–2345, 2010

    2010

  28. [28]

    Big data in a nano world: a review on computational, data-driven design of nanomaterials structures, prop- erties, and synthesis.ACS Nano, 16(12):19873–19891, 2022

    Ruo Xi Yang, Caitlin A McCandler, Oxana Andriuc, Martin Siron, Rachel Woods- Robinson, Matthew K Horton, and Kristin A Persson. Big data in a nano world: a review on computational, data-driven design of nanomaterials structures, prop- erties, and synthesis.ACS Nano, 16(12):19873–19891, 2022

    2022

  29. [29]

    Equivariant diffusion for crystal structure prediction

    Peijia Lin, Pin Chen, Rui Jiao, Qing Mo, Jianhuan Cen, Wenbing Huang, Yang Liu, Dan Huang, and Yutong Lu. Equivariant diffusion for crystal structure prediction. arXiv preprint arXiv:2512.07289, 2025

    arXiv 2025

  30. [30]

    Space group equivariant crystal diffusion.arXiv preprint arXiv:2505.10994, 2025

    Rees Chang, Angela Pak, Alex Guerra, Ni Zhan, Nick Richardson, Elif Ertekin, and Ryan P Adams. Space group equivariant crystal diffusion.arXiv preprint arXiv:2505.10994, 2025

    arXiv 2025

  31. [31]

    Symmetry-informed geometric representation for molecules, proteins, and crystalline materials.Advances in neural information processing systems, 36:66084–66101, 2023

    Shengchao Liu, Yanjing Li, Zhuoxinran Li, Zhiling Zheng, Chenru Duan, Zhi- Ming Ma, Omar Yaghi, Animashree Anandkumar, Christian Borgs, Jennifer Chayes, et al. Symmetry-informed geometric representation for molecules, proteins, and crystalline materials.Advances in neural information processing systems, 36:66084–66101, 2023

    2023

  32. [32]

    Kohn-sham density functional theory: predicting and understanding chemistry.Reviews in Computational Chemistry, pages 1–86, 2000

    F Matthias Bickelhaupt and Evert Jan Baerends. Kohn-sham density functional theory: predicting and understanding chemistry.Reviews in Computational Chemistry, pages 1–86, 2000

    2000

  33. [33]

    Perspective: Kohn-sham density functional theory descending a staircase.The Journal of Chemical Physics, 145(13), 2016

    Haoyu S Yu, Shaohong L Li, and Donald G Truhlar. Perspective: Kohn-sham density functional theory descending a staircase.The Journal of Chemical Physics, 145(13), 2016

    2016

  34. [34]

    Sparsity of the density matrix in kohn- sham density functional theory and an assessment of linear system-size scaling methods.Physical Review Letters, 79(20):3962, 1997

    Roi Baer and Martin Head-Gordon. Sparsity of the density matrix in kohn- sham density functional theory and an assessment of linear system-size scaling methods.Physical Review Letters, 79(20):3962, 1997

    1997

  35. [35]

    Introducing onetep: Linear-scaling density functional simulations on parallel computers.The Journal of Chemical Physics, 122(8), 2005

    Chris-Kriton Skylaris, Peter D Haynes, Arash A Mostofi, and Mike C Payne. Introducing onetep: Linear-scaling density functional simulations on parallel computers.The Journal of Chemical Physics, 122(8), 2005

    2005

  36. [36]

    Guishan Zheng, Stephan Irle, and Keiji Morokuma. Performance of the dftb method in comparison to dft and semiempirical methods for geometries and energies of c20–c86 fullerene isomers.Chemical Physics Letters, 412(1-3):210–216, 2005

    2005

  37. [37]

    Density-functional tight-binding: basic concepts and applications to molecules and clusters.Ad- vances in Physics: X, 5(1):1710252, 2020

    Fernand Spiegelman, Nathalie Tarrat, Jérôme Cuny, Leo Dontot, Evgeny Posen- itskiy, Carles Martí, Aude Simon, and Mathias Rapacioli. Density-functional tight-binding: basic concepts and applications to molecules and clusters.Ad- vances in Physics: X, 5(1):1710252, 2020

    2020

  38. [38]

    Analytical approach to phonon calculations in the scc-dftb framework.The Journal of Chemical Physics, 153(14), 2020

    Vladimir Bačić, Thomas Heine, and Agnieszka Kuc. Analytical approach to phonon calculations in the scc-dftb framework.The Journal of Chemical Physics, 153(14), 2020

    2020

  39. [39]

    Multiscale modeling of agglomerated ceria nanoparticles: interface stability and oxygen vacancy formation.Frontiers in Chemistry, 7:203, 2019

    Byung-Hyun Kim, Jolla Kullgren, Matthew J Wolf, Kersti Hermansson, and Peter Broqvist. Multiscale modeling of agglomerated ceria nanoparticles: interface stability and oxygen vacancy formation.Frontiers in Chemistry, 7:203, 2019

    2019

  40. [40]

    Embedded-atom method: Derivation and application to impurities, surfaces, and other defects in metals.Physical Review B, 29(12):6443, 1984

    Murray S Daw and Michael I Baskes. Embedded-atom method: Derivation and application to impurities, surfaces, and other defects in metals.Physical Review B, 29(12):6443, 1984

    1984

  41. [41]

    Generalized neural-network representa- tion of high-dimensional potential-energy surfaces.Physical Review Letters, 98(14):146401, 2007

    Jörg Behler and Michele Parrinello. Generalized neural-network representa- tion of high-dimensional potential-energy surfaces.Physical Review Letters, 98(14):146401, 2007

    2007

  42. [42]

    Avik Mahata, Tanmoy Mukhopadhyay, and Mohsen Asle Zaeem. Modified embedded-atom method interatomic potentials for al-cu, al-fe and al-ni binary alloys: From room temperature to melting point.Computational Materials Science, 201:110902, 2022

    2022

  43. [43]

    Enhancing the electronic properties of tio2 nanoparticles through carbon doping: An integrated dftb and computer vision approach.Computational Materials Science, 244:113248, 2024

    Mustafa Kurban, Can Polat, Erchin Serpedin, and Hasan Kurban. Enhancing the electronic properties of tio2 nanoparticles through carbon doping: An integrated dftb and computer vision approach.Computational Materials Science, 244:113248, 2024

    2024

  44. [44]

    L. C. Blum and J.-L. Reymond. 970 million druglike small molecules for virtual screening in the chemical universe database GDB-13.J. Am. Chem. Soc., 131:8732, 2009

    2009

  45. [45]

    M. Rupp, A. Tkatchenko, K.-R. Müller, and O. A. von Lilienfeld. Fast and accurate modeling of molecular atomization energies with machine learning.Physical Review Letters, 108:058301, 2012

    2012

  46. [46]

    Accurate global machine learning force fields for molecules with hundreds of atoms.Science Advances, 9(2):eadf0873, 2023

    Stefan Chmiela, Valentin Vassilev-Galindo, Oliver T Unke, Adil Kabylda, Huziel E Sauceda, Alexandre Tkatchenko, and Klaus-Robert Müller. Accurate global machine learning force fields for molecules with hundreds of atoms.Science Advances, 9(2):eadf0873, 2023

    2023

  47. [47]

    Pubchem 2025 update.Nucleic Acids Research, 53(D1):D1516–D1525, 2025

    Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li, Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, et al. Pubchem 2025 update.Nucleic Acids Research, 53(D1):D1516–D1525, 2025

    2025

  48. [48]

    nabladft: Large-scale conformational energy and hamil- tonian prediction benchmark and dataset.Physical Chemistry Chemical Physics, 24(42):25853–25863, 2022

    Kuzma Khrabrov, Ilya Shenbin, Alexander Ryabov, Artem Tsypin, Alexander Telepov, Anton Alekseev, Alexander Grishin, Pavel Strashnov, Petr Zhilyaev, Sergey Nikolenko, et al. nabladft: Large-scale conformational energy and hamil- tonian prediction benchmark and dataset.Physical Chemistry Chemical Physics, 24(42):25853–25863, 2022

    2022

  49. [49]

    Qh9: A quantum hamiltonian prediction benchmark for qm9 molecules.Advances in Neural Information Processing Systems, 36, 2024

    Haiyang Yu, Meng Liu, Youzhi Luo, Alex Strasser, Xiaofeng Qian, Xiaoning Qian, and Shuiwang Ji. Qh9: A quantum hamiltonian prediction benchmark for qm9 molecules.Advances in Neural Information Processing Systems, 36, 2024

    2024

  50. [50]

    New cubic perovskites for one-and two-photon water splitting using the computational materials repository

    Ivano E Castelli, David D Landis, Kristian S Thygesen, Søren Dahl, Ib Chork- endorff, Thomas F Jaramillo, and Karsten W Jacobsen. New cubic perovskites for one-and two-photon water splitting using the computational materials repository. Energy & Environmental Science, 5(10):9034–9043, 2012

    2012

  51. [51]

    Computational screening of perovskite metal oxides for optimal solar light capture.Energy & Environmental Science, 5(2):5814–5819, 2012

    Ivano E Castelli, Thomas Olsen, Soumendu Datta, David D Landis, Søren Dahl, Kristian S Thygesen, and Karsten W Jacobsen. Computational screening of perovskite metal oxides for optimal solar light capture.Energy & Environmental Science, 5(2):5814–5819, 2012

    2012

  52. [52]

    Airss data for carbon at 10gpa and the c+ n+ h+ o system at 1gpa.(No Title), 2020

    Chris J Pickard. Airss data for carbon at 10gpa and the c+ n+ h+ o system at 1gpa.(No Title), 2020

    2020

  53. [53]

    Bench- marking materials property prediction methods: the matbench test set and au- tomatminer reference algorithm.npj Computational Materials, 6(1):138, 2020

    Alexander Dunn, Qi Wang, Alex Ganose, Daniel Dopp, and Anubhav Jain. Bench- marking materials property prediction methods: the matbench test set and au- tomatminer reference algorithm.npj Computational Materials, 6(1):138, 2020

    2020

  54. [54]

    The open catalyst 2022 (oc22) dataset and challenges for oxide electrocatalysts.ACS Catalysis, 13(5):3066–3084, 2023

    Richard Tran, Janice Lan, Muhammed Shuaibi, Brandon M Wood, Siddharth Goyal, Abhishek Das, Javier Heras-Domingo, Adeesh Kolluru, Ammar Rizvi, Nima Shoghi, et al. The open catalyst 2022 (oc22) dataset and challenges for oxide electrocatalysts.ACS Catalysis, 13(5):3066–3084, 2023. 9

    2022

  55. [55]

    Lambench: A benchmark for large atomic models.arXiv preprint arXiv:2504.19578, 2025

    Anyang Peng, Chun Cai, Mingyu Guo, Duo Zhang, Chengqian Zhang, Antoine Loew, Linfeng Zhang, and Han Wang. Lambench: A benchmark for large atomic models.arXiv preprint arXiv:2504.19578, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    Crysmtm: a multiphase, temperature-resolved, multimodal dataset for crystalline materials

    Can Polat, Erchin Serpedin, Mustafa Kurban, and Hasan Kurban. Crysmtm: a multiphase, temperature-resolved, multimodal dataset for crystalline materials. Machine Learning: Science and Technology, 6(3):030603, 2025

    2025

  57. [57]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    Pith/arXiv arXiv 2001

  58. [58]

    Scaling theory for finite-size effects in the critical region.Physical Review Letters, 28(23):1516, 1972

    Michael E Fisher and Michael N Barber. Scaling theory for finite-size effects in the critical region.Physical Review Letters, 28(23):1516, 1972

    1972

  59. [59]

    C2np: A bench- mark for learning scale-dependent geometric invariances in 3d materials genera- tion.arXiv preprint arXiv:2601.19076, 2026

    Can Polat, Erchin Serpedin, Mustafa Kurban, and Hasan Kurban. C2np: A bench- mark for learning scale-dependent geometric invariances in 3d materials genera- tion.arXiv preprint arXiv:2601.19076, 2026

    arXiv 2026

  60. [60]

    H. W. King.CRC Handbook of Chemistry and Physics. CRC Press, 83 edition, 2002. Standard phase data for silver (Ag)

    2002

  61. [61]

    H. W. King.CRC Handbook of Chemistry and Physics. CRC Press, 83 edition, 2002. Standard phase data for gold (Au)

    2002

  62. [62]

    L. W. Finger and R. M. Hazen. Crystal structure and isothermal compression of fe2o3, cr2o3, and v2o3 to 50 kbars.Journal of Applied Physics, 51:5362–5367, 1980

    1980

  63. [64]

    Grau-Crespo and R

    R. Grau-Crespo and R. Lopez-Cordero. Mos 2 structural properties.Phys. Chem. Chem. Phys., 4:4078, 2002

    2002

  64. [66]

    W. H. Baur, R. A. Sass, et al. The rutile structure of sno2.Acta Crystallographica Section B, 27:2133, 1971

    1971

  65. [67]

    Crystal chemistry of perovskite-type compounds in the tausonite-loparite series,(sr1- 2 x na x la x) tio3.Physics and Chemistry of Minerals, 27(8):583–589, 2000

    Roger H Mitchell, Anton R Chakhmouradian, and Patrick M Woodward. Crystal chemistry of perovskite-type compounds in the tausonite-loparite series,(sr1- 2 x na x la x) tio3.Physics and Chemistry of Minerals, 27(8):583–589, 2000

    2000

  66. [68]

    M. Horn, C. R. Meagher, et al. Structure of anatase tio2.Zeitschrift für Kristallo- graphie, 136:273, 1972

    1972

  67. [69]

    R. W. G. Wyckoff.Crystal Structures Volume 1. Interscience Publishers, 1963

    1963

  68. [70]

    Wmd-group/hybrid- perovskites: Collection 1 (v1.0)

    Aron Walsh, elds22, Federico Brivio, and Jarvist Moore Frost. Wmd-group/hybrid- perovskites: Collection 1 (v1.0). https://doi.org/10.5281/zenodo.2641358, 2019. Hybrid perovskite CH3NH3PbI3 structural data

    work page doi:10.5281/zenodo.2641358 2019

  69. [71]

    Crystal diffusion variational autoencoder for periodic material generation.arXiv preprint arXiv:2110.06197, 2021

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola. Crystal diffusion variational autoencoder for periodic material generation.arXiv preprint arXiv:2110.06197, 2021

    Pith/arXiv arXiv 2021

  70. [72]

    Crystal structure prediction by joint equivariant diffusion.Advances in Neural Information Processing Systems, 36:17464–17497, 2023

    Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. Crystal structure prediction by joint equivariant diffusion.Advances in Neural Information Processing Systems, 36:17464–17497, 2023

    2023

  71. [73]

    Flowmm: Generating materials with riemannian flow matching

    Benjamin Kurt Miller, Ricky TQ Chen, Anuroop Sriram, and Brandon M Wood. Flowmm: Generating materials with riemannian flow matching. InForty-first International Conference on Machine Learning, 2024

    2024

  72. [74]

    Mattergen: a generative model for inorganic materials design.arXiv preprint arXiv:2312.03687, 2023

    Claudio Zeni, Robert Pinsler, Daniel Zügner, Andrew Fowler, Matthew Hor- ton, Xiang Fu, Sasha Shysheya, Jonathan Crabbé, Lixin Sun, Jake Smith, et al. Mattergen: a generative model for inorganic materials design.arXiv preprint arXiv:2312.03687, 2023

    Pith/arXiv arXiv 2023

  73. [75]

    All-atom diffusion transform- ers: Unified generative modelling of molecules and materials.arXiv preprint arXiv:2503.03965, 2025

    Chaitanya K Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary W Ulissi. All-atom diffusion transform- ers: Unified generative modelling of molecules and materials.arXiv preprint arXiv:2503.03965, 2025. 10

    Pith/arXiv arXiv 2025