Pith. sign in

REVIEW 4 major objections 6 minor 20 references

Topological representation of layered hybrid lead halides for machine-learning using universal clusters

T0 review · 4 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Atom-specific distance barcodes make layered lead-halide structures machine-readable and predict DFT band gaps with $R^2 = 0.8$ and $\mathrm{MAE} = 0.12$ eV on 140 $n=1$ (100) compounds.

desk verdict The paper is a plausible barcode-based band-gap model for 2D lead halides, but the reported metrics are confounded by composition features and need an ablation. read the letter →

arxiv 2411.11122 v1 pith:DKZRYOMA submitted 2024-11-17 cond-mat.mtrl-sci

Ekaterina I. Marchenko , Maria G. Khrenova , Korolev V.V. , Eugene A. Goodilin , Alexey B. Tarasov This is my paper · ORCID
classification cond-mat.mtrl-sci
keywords hybridhalideperovskitestopologicalrepresentationspersistenthomologybandgapsstructure-propertyrelationshipsmachinelearninglayeredleadhalidesgradientboosting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper seeks to establish that the geometry of a layered hybrid lead-halide crystal can be encoded as a set of atom-centered distance barcodes, and that this representation alone, together with elemental composition statistics, is enough for a machine-learning model to predict band gaps as accurately as contemporary methods. Using 140 two-dimensional (100) perovskite structures with perovskite block thickness $n=1$, a gradient-boosted tree model reaches $R^2 = 0.8$, $\mathrm{RMSE} = 0.17$ eV, and $\mathrm{MAE} = 0.12$ eV against DFT reference gaps. If this holds, band-gap screening for photovoltaic and optoelectronic lead halides could run directly from crystallographic files without hand-picked structural descriptors, and the same topological representation could extend to 3D, 1D, and 0D hybrid halides.

What carries the argument

The carrying mechanism is the atom-specific persistent homology barcode, a topological fingerprint computed around each crystallographically distinct atom in the unit cell. A sphere with a cutoff radius of 10 Å is placed on the atom, and every atomic neighbor inside the sphere produces a line segment whose length equals the interatomic distance; the number of overlapping segments reports the multiplicity of that contact in the unit cell. Because every geometric parameter is reduced to the same homogeneous object, an interatomic distance, the representation is invariant and needs no hand-chosen unit-cell parameters, angles, or coordinates. The barcode is augmented with composition-based statistics, and the analysis singles out the Pb-X, X-X, and N-X atom pairs (X = halogen) as the channels whose distance changes most strongly affect band gaps. A gradient-boosted regression tree maps the resulting feature vectors to DFT band gaps.

What would settle it

Train the same barcode-plus-gradient-boosted pipeline on only one halide chemistry, say lead iodides, and test it on lead bromides or lead chlorides with DFT band gaps; if the mean absolute error degrades far beyond 0.12 eV, the universality and cutoff claims would be contradicted.

Watch

Extended reading notes

Core claim

The central claim is that atom-specific persistent homology barcodes provide a universal and invariant machine-readable representation of layered hybrid lead halide structures. Around each atom in the unit cell, a 10 Å sphere is drawn; every neighbor inside contributes a line to the barcode, with the line length equal to the interatomic distance and the multiplicity recording how often that contact appears in the cell. A gradient-boosted tree trained on these barcodes plus composition statistics predicts DFT band gaps of 140 $n=1$ (100) structures with $R^2 = 0.8$, $\mathrm{RMSE} = 0.17$ eV, and $\mathrm{MAE} = 0.12$ eV. The paper further argues that the same representation can support both the direct problem of predicting properties from structure and the inverse problem of proposing structures with desired properties, extending to other hybrid halide dimensionalities.

Load-bearing premise

The central claim rests on the assumptions that a 10 Å cutoff around each atom captures all band-gap-relevant structural information and that the 140 $n=1$ (100) structures in the training database are representative enough for cross-validated accuracy to transfer to other layered hybrid lead halides.

Editorial extensions

If this is right

  • Band gaps of $n=1$ (100) layered hybrid lead halides can be estimated from a crystallographic information file alone, with a reported mean absolute error of 0.12 eV, which is enough for rapid compositional screening.
  • Because no hand-crafted geometric descriptors are required, the same barcode pipeline can in principle be applied to any periodic crystal structure without changing the feature-engineering step.
  • The atom-pair analysis indicates that barcodes built from Pb-X, X-X, and N-X contacts carry most of the band-gap signal, so future descriptor design can concentrate resolution on these channels.
  • The paper positions the barcode representation for both direct structure-to-property prediction and inverse property-to-structure design, and states that the approach can extend to 3D, 1D, and 0D hybrid halide materials.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural stress test of the universality claim is cross-anion transfer: train the same barcode-plus-tree pipeline on iodide-only structures and evaluate on bromide and chloride compounds; the 10 Å cutoff and the Pb-X/X-X/N-X channels would need to generalize for the method to hold beyond the 140-compound training set.
  • The reported $R^2 = 0.8$ is likely helped by the homogeneity of the $n=1$ (100) family; applying the pipeline to a mixed set of $n=1$, $n=2$, and 3D lead halides would reveal whether the topological representation alone spans different inorganic substructures without retraining.
  • Because a barcode is essentially a distance multiset around each atom, it constrains candidate geometries in principle, so the paper's inverse-design goal could be pursued by decoding barcodes into coordinates; the paper does not yet demonstrate that decoding step.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes the use of atom-specific persistent homology barcodes, following ref. 13, as a machine-readable representation for layered hybrid lead halides (LHPs). For each atom in the unit cell, a spherical local environment with a 10 Å cutoff generates a barcode encoding interatomic distances. These barcodes are then combined with composition-based features (elemental fractions, element-property statistics, and electronic structure attributes) and fed into a gradient-boosted regression tree (GBRT) to predict DFT band gaps for 140 n=1 (100) LHP structures from the authors' NMSE database. The authors report tenfold cross-validation metrics of R2=0.8, RMSE=0.17 eV, and MAE=0.12 eV and conclude that the topological representation is a universal, invariant, and suitable general-purpose representation for this class of materials.

Significance. The paper's core idea—using persistent homology barcodes as invariant structural descriptors for layered halide perovskites—is timely and potentially useful for structure-based screening. The repeated tenfold CV protocol is a reasonable internal validation scheme, and the dataset and underlying barcode construction method are traceable to published references. However, the significance of the work is currently limited because the headline metrics are obtained from a feature vector that concatenates barcodes with composition-based descriptors, and the manuscript provides no ablation or compositional baseline. The potential contribution of the barcodes to the predictive accuracy is therefore unquantified, and the claim of a 'good general-purpose machine-readable representation' is not yet supported.

major comments (4)
  1. [Methods paragraph after Figure 1 description] The paragraph immediately following the barcode description states that 'composition-based features were incorporated, which include stoichiometric attributes reflecting elemental fractions, elemental-property statistics... and electronic structure attributes.' The reported R2=0.8, RMSE=0.17 eV, and MAE=0.12 eV are for this combined feature vector. Because the 140 structures vary in composition and stoichiometry, a GBRT trained only on the composition features could plausibly achieve comparable accuracy without using any structural information. The paper should report results for (i) barcodes only, (ii) composition features only, and (iii) the combined vector under the identical repeated tenfold CV protocol, together with feature-importance analysis. Additionally, the term 'electronic structure attributes' is undefined; the authors should specify exactly what these features are and confirm that none are derived from the target band gaps. Without this ablation, the central claim that the topological barcode representation is responsible for the predictive accuracy is not established.
  2. [Hyperparameter footnote iii] The text states 'Voronoi tessellations and Coulomb matrices were replicated using Magpie' but no comparison metrics are presented. Since the paper claims an advantage over structure-graph and Coulomb-matrix representations, a quantitative comparison on the same data and CV splits is necessary. Please add a table reporting the same error metrics for the Magpie descriptors under the identical validation setup.
  3. [Dataset description] The 140 structures are exclusively n=1 (100) layered hybrid lead halides from the authors' own NMSE database (ref. 8), and all reported metrics come from internal tenfold CV. Because the structural variety is limited, the claim that the representation is universal and transferable to other hybrid halide families (e.g., 3D, 1D, 0D substructures, or n>1) requires at least a holdout test on a few structurally distinct compounds not seen during training, or a leave-one-family-out cross-validation. Without such an external check, the 'universal' claim remains a conjecture.
  4. [Footnote ii] Footnote ii asserts that 'a cutoff radius of 10 Å ... is optimal' for LHP structures because interlayer interactions are possible up to such distances. No sensitivity analysis is provided. Since the barcodes are defined entirely by the local environment within the cutoff, the paper should report model performance for at least two alternative cutoffs (e.g., 6 Å and 12 Å) to show that the 10 Å choice is not a critical hyperparameter. If the performance varies strongly, the optimality claim needs an explicit justification.
minor comments (6)
  1. [Paragraph after Figure 1] The sentence 'The order of the lines in the barcode does can be chosen arbitrarily.' contains a typo; it should read 'does not' or 'can be'.
  2. [Dataset description] The phrase 'This dataset i was sourced from 8' contains a stray lowercase 'i'; please rephrase.
  3. [Abstract / text] One instance of 'Сrystallographic' uses a Cyrillic 'С'; replace with the Latin 'C'.
  4. [Data availability] No data or code availability statement is given; please provide access to the barcode construction scripts and feature generation code to allow reproducibility.
  5. [Figure 3] Figure 3 does not include the standard deviation of the repeated CV runs; please add error bars or report the spread in the caption or a table.
  6. [References] The reference list contains formatting inconsistencies (e.g., ref. 17 journal volume formatting); please normalize according to journal style.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the band-gap prediction is a standard cross-validated regression on structural and composition features; the main weakness is a confounding/ablation gap, not a self-referential derivation.

full rationale

The paper's derivation chain is: CIF structures -> atom-centered 10 Å point clouds -> persistent-homology barcodes -> feature vector (barcodes plus composition features) -> GBRT -> predicted band gaps. Predicted band gaps are compared with DFT labels from the authors' prior database (ref. 8), but this is data provenance, not definitional circularity: no equation introduces the DFT band gap as an input to the model, and the reported R2=0.8/RMSE=0.17 eV/MAE=0.12 eV come from tenfold cross-validation repeated 20 times, so the 'prediction' is an honest out-of-sample holdout estimate rather than a fitted quantity renamed as a prediction. The 10 Å cutoff is asserted as optimal (footnote ii), but this is an unproven modeling assumption, not a circular step; the same is true for the absence of an ablation separating topological barcodes from composition-based features. That absence weakens the causal claim that barcodes drive accuracy, but a confounding/evidence gap is not equivalence-by-construction or a self-citation chain. Self-citations (refs 6, 8, 11, 12, 17) supply the dataset and prior structure-property context, but the central algorithm is attributed to the external ref. 13, and no load-bearing step reduces to an unverified citation by the same authors. Therefore no enumerated circularity pattern can be exhibited with the required quote-and-reduction specificity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the accuracy of the DFT labels, the correctness of the externally developed barcode method, the sufficiency of the hand-set cutoff, and the representativeness of the dataset. No new physical entities are introduced. Free parameters are limited to the cutoff radius and the ML hyperparameters, both selected using the same data family on which the model is evaluated.

free parameters (2)
  • cutoff radius for local atomic environment = 10 Å
    Chosen by hand and described as optimal for LHP because interactions between adjacent layers occur up to such distances; the universality claim depends on this choice.
  • GBRT hyperparameters = n_estimators=300000, learning_rate=0.001, max_depth=7, min_samples_split=5, subsample=0.85, max_features=sqrt
    Optimized via cross-validation on the same dataset; they influence the reported metrics but are standard model parameters rather than physical constants.
assumptions (4)
  • domain assumption DFT band gaps from the NMSE database (Marchenko et al. 2020) are accurate ground truths.
    Used as labels for training and validation; no experimental verification or error estimates are provided in this paper.
  • standard math The barcode construction of Jiang et al. (2021) is correctly implemented and is invariant as claimed.
    The method is adopted from reference 13, not re-derived or verified here.
  • ad hoc to paper A cutoff radius of 10 Å captures all relevant structural information for band gaps of 2D LHPs.
    Stated as optimal without a systematic convergence study; band gap may depend on longer-range interactions.
  • domain assumption The selected structures are representative of all layered hybrid lead halides with n=1 (100) type.
    All 140 structures come from one database and no independent test set is used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Topological representation of layered hybrid lead halides for machine-learning using universal clusters." pith.science (2026). https://pith.science/paper/DKZRYOMA

@misc{pith2026241111122,
  author       = {Pith},
  title        = {Pith review of: Topological representation of layered hybrid lead halides for machine-learning using universal clusters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKZRYOMA}},
  note         = {Machine review of arXiv:2411.11122}
}
read the original abstract

Layered hybrid halide compounds offer promising functional properties, particularly tunable band gaps, conductivity, light harvesting thus making them prospective for applications in photovoltaics and optoelectronics. This study exemplifies an approach of predicting band gaps using machine learning models enhanced by invariant topological representations of these materials using the atom-specific persistent homology method in order to facilitate the discovery and design of new hybrid halide materials with tailored electronic properties.

Figures

Figures reproduced from arXiv: 2411.11122 by the authors.

Figure 1
Figure 1. Illustration of barcodes construction of LHP crystal structure. The crystal structures were visualized using the Vesta15 and TOPOSpro16 programs. The methodology and algorithm for constructing specific barcodes in Python were developed as outlined in reference 13. The fundamental concept underlying this approach is that ii To create a region of point clouds surrounding each atom, we employed a cutoff radius of 10 Å … view at source ↗
Figure 3
Figure 3. Comparison of DFT-calculated band gaps and predicted band gaps by ML algorithm for 2D hybrid lead halide materials. Beyond LHP materials, this approach presents opportunities for addressing both the direct problem (predicting the physical properties of materials from their crystal structure) and the inverse problem (predicting crystal structures with desired properties) for other hybrid materials related to the grou… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 20 canonical work pages

  1. [1]

    Blancon, J

    J.C. Blancon, J. Even, C.C. Stoumpos, M.G. Kanatzidis, and A.D. Mohite. Nat. Nanotechnol., 2020, 1512, 15, pp. 969–985

    work page 2020

  2. [2]

    Huang, Y

    J. Huang, Y. Yuan, Y. Shao, and Y. Yan. Nat. Rev. Mater., 2017, 27, 2, pp. 1–19

    work page 2017

  3. [3]

    W. Li, Z. Wang, F. Deschler, S. Gao, R.H. Friend and A.K. Cheetham. Nat. Rev. Mater., 2017, 23, 2, pp. 1–18

    work page 2017

  4. [4]

    Grancini, M.K

    G. Grancini, M.K. Nazeeruddin. Nat. Rev. Mater., 2019, 4, pp. 4–22

    work page 2019

  5. [5]

    Mao, C.C

    L. Mao, C.C. Stoumpos, M.G. Kanatzidis. J. Am. Chem. Soc., 2019, 141, pp.1171–1190

    work page 2019

  6. [6]

    Marchenko, V.V

    E.I. Marchenko, V.V. Korolev, S.A. Fateev, A. Mitrofanov, N.N. Eremin, E.A. Goodilin, and A.B. Tarasov. Chem. Mater., 2021, 33, pp. 7518–7526

    work page 2021

  7. [7]

    Z. Wan, Q. De Wang, D. Liu, J. Liang. New J. Chem., 2021, 45, pp. 9427–9433

    work page 2021

  8. [8]

    Marchenko, S.A

    E.I. Marchenko, S.A. Fateev, A.A. Petrov, V.V. Korolev, A. Mitrofanov, A.V. Petrov, E.A. Goodilin and A.B. Tarasov. Chem. Mater., 2020, 32, pp. 7383–7388

    work page 2020

Show all 20 references
  1. [9]

    W.B. Park, J. Chung, J. Jung, K. Sohn, S. P. Singh, M. Pyo, N. Shin, K.-S. Sohn. IUCrJ , 2017, 4, pp. 486–494

    2017

  2. [10]

    C.S. Hu, R. Mayengbam, M.C. Wu, K. Xia, T.C. Sum. Commun. Mater., 2024, 5, pp. 1–10

    2024

  3. [11]

    Marchenko, V.V

    E.I. Marchenko, V.V. korolev, A. Mitrofanov, S.A. Fateev, E.A. Goodilin, A.B. Tarasov. Chem. Mater., 2021, 33, pp. 1213–1217

    2021

  4. [12]

    Marchenko, S.A

    E.I. Marchenko, S.A. Fateev, A.A. Ordinartsev, P.A. Ivlev, E.A. Goodilin, A.B. Tarasov. Mendeleev Commun., 2022, 32, pp. 315–316

    2022

  5. [13]

    Jiang, D

    Y. Jiang, D. Chen, X. Chen, T. Li, G.-W. Wei, F. Pan. npj Comput. Mater., 2021, 7, pp. 1–8

    2021

  6. [14]

    Meredig, A

    B. Meredig, A. Agrawal, S. Kirklin, J.E. Saal, J.W. Doak, A. Thompson, K. Zhang, A. Choudhary, and C. Wolverton. Phys. Rev. B - Condens. Matter Mater. Phys., 2014, 89, pp. 1–7

    2014

  7. [15]

    Momma, F

    K. Momma, F. Izumi. J. Appl. Crystallogr., 2011, 44, pp. 1272–1276

    2011

  8. [16]

    Blatov, A.P

    V.A. Blatov, A.P. Shevchenko, D.M. Proserpio. Cryst. Growth Des., 2014, 14, pp. 3576–3586

    2014

  9. [17]

    Marchenko, S.A

    E.I. Marchenko, S.A. fateev, V.V. Korolev, V. Buchinskiy, N.N. Eremin, E.A. Goodilin, and A.B. Tarasov. J. Mater. Chem. C, 2022, 44, 10, pp. 16838-16846

    2022

  10. [18]

    Li, Y Liu, D

    S. Li, Y Liu, D. Chen, Y. Jiang, Z. Nie, F. Pan. Wiley Interdiscip. Rev. Comput. Mol. Sci. , 2022, 12, pp. 1–20

    2022

  11. [19]

    Pedregosa, G

    F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, É. Duchesnay. J. Mach. Learn. Res., 2011, 12, pp. 2825–2830

    2011

  12. [20]

    L. Ward, A. Agrawal, A. Choudhary, C. Wolverton. npj Comput. Mater., 2016, 2, pp. 1–7. Graphical abstract

    2016

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.