Quotient Complex Transformer (QCformer) for Perovskite Data Analysis

Chuan-Shen Hu; Kelin Xia; Tze Chien Sum; Xiang Liu; Xinyu You

REVIEW 4 major objections 4 minor 79 references

Quotient Complex Transformer (QCformer) for Perovskite Data Analysis

T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A crystal becomes a quotient complex—edges for bonds, triangles for three-atom interactions—and a simplex Transformer over it predicts perovskite bandgaps more accurately than graph networks.

desk verdict A novel and promising quotient-complex transformer for crystal bandgap prediction, with the triangle-feature story undermined by degenerate self-loop simplices and a lack of error bars. read the letter →

arxiv 2505.09174 v1 pith:TEYLAKMC submitted 2025-05-14 cs.LG cond-mat.mtrl-sci

Xinyu You , Xiang Liu , Chuan-Shen Hu , Kelin Xia , Tze Chien Sum This is my paper · ORCID

classification cs.LGcond-mat.mtrl-sci

keywords quotientcomplexsimplicialhigher-orderinteractionsperovskitebandgappredictiontopologicaldeeplearninggeometricmaterialpropertytransfer

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

QCformer models a crystal as a quotient complex: periodically equivalent atoms are glued together into one finite cell, pairwise bonds become edges, and three-atom interactions become triangles. The paper's central claim is that this representation, processed by a Transformer that passes messages from triangles to edges to vertices, captures periodic structure and higher-order interactions that ordinary graph neural networks miss. On the HOIP perovskite bandgap benchmark it reports a 30.59% reduction in MAE over MEGNet (0.0969 eV versus 0.1396 eV), with a coefficient of determination of 0.9829 and a Pearson correlation coefficient of 0.9916, and on the HOIP2D benchmark it lowers MAE by about 9.2% relative to the best previous GNN while improving RMSE. An ablation that removes triangle features, leaving only the quotient graph, degrades performance on every task tested, which the paper presents as evidence that the three-body features are what carry the improvement. If correct, the quotient complex gives a general way to embed periodicity and many-body physics into deep-learning models for materials.

What carries the argument

The load-bearing object is the quotient complex. For a given crystal, the paper builds a directed 12-nearest-neighbor graph on the unit cell's atoms, forms its clique complex so that mutually connected triples become 2-simplices (triangles), and then identifies all atoms related by lattice translations to obtain a finite quotient complex with multi-edges, self-loops, and triangles. Triangle features are built from the three edge lengths, their pairwise products, and their squares, expanded through radial basis functions. The Simplex Transformer (Sformer) block updates an n-simplex by attending from its neighboring n-simplices through shared (n+1)-simplex cofaces; QCformer runs five edge-to-vertex layers followed by two triangle-to-edge-to-vertex layers. The mathematical analysis proves that the quotient complex is homotopy equivalent to an auxiliary simplicial complex K~ obtained by adding star-shaped sets, so the homology groups are preserved; this is how the paper argues the periodic topology of the crystal survives the quotient.

What would settle it

Retrain QCformer on the HOIP benchmark twice: once with all 2-simplices that contain a self-loop edge removed, and once with triangle features replaced by features of randomly selected triples of edges of the same size and dimension. If removing self-loop triangles leaves MAE essentially unchanged, or if random triples reproduce the 0.0969 eV MAE, then the claimed three-body encoding is not what drives the improvement.

Watch

Extended reading notes

Core claim

The paper's core claim is that representing a periodic crystal as a quotient complex rather than a plain graph improves material property prediction. Starting from a 12-nearest-neighbor graph around each atom in the unit cell, the construction takes its clique complex and glues together atoms related by lattice translations, yielding vertices, multi-edges, self-loops, and triangles; triangles are assigned features from the three edge lengths and their products and squares. A Simplex Transformer then updates simplex embeddings through attention over cofaces, with the hierarchy flowing from triangles to edges to vertices. On the HOIP dataset the model reaches MAE 0.0969 eV, a 30.59% improvement over MEGNet's 0.1396 eV, with COD 0.9829 and PCC 0.9916; on HOIP2D it reaches MAE 0.0754 eV and RMSE 0.1038 eV, lower than the strongest prior GNN SIGNNA_c and competitive with descriptor-based gradient-boosting models. On the Materials Project and JARVIS benchmarks it posts the best or tied-best MAE on most tasks, including a 13.16% improvement over the second-best model on Bulk Moduli and 7.38% and 7.69% improvements on the two JARVIS bandgap tasks, and removing triangle features in the 1D-QCformer ablation hurts every task.

Load-bearing premise

The method's advantage rests on the assumption that the triangles formed by gluing periodic copies of the 12-nearest-neighbor graph encode real three-atom chemical interactions, even when a triangle contains a self-loop edge rather than three ordinary bonds.

Editorial extensions

If this is right

Pretrained on large inorganic datasets (Materials Project and JARVIS) and fine-tuned on small perovskite datasets, the model transfers well enough to give HOIP bandgap MAE under 0.1 eV from only 1346 data points.
Because the graph-only ablation is worse on all nine benchmark tasks, the triangle-to-edge message flow, not just the Transformer backbone, is the source of the gain; adding tetrahedral 3-simplices is the natural next extension.
On five unlabeled 2D perovskites, QCformer's predicted bandgaps fall between GGA-PBE and r2SCAN DFT values in most cases, suggesting it can serve as a cheap screening estimator when experimental labels do not exist.
The same quotient-complex representation should apply to any periodic solid, not just perovskites, wherever many-body interactions matter.

Reading between the lines

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

The paper does not test whether triangles containing a self-loop edge after the quotient gluing are physically meaningful; deleting exactly those degenerate 2-simplices and re-running the HOIP experiment would isolate whether the 30.59% gain comes from three-body chemistry or from extra message-passing capacity.
If the gain survives randomly permuted triangle features, then the advantage may be architectural rather than physical; the same parameter budget spent on extra edge-to-vertex attention layers could be compared against QCformer to separate the two explanations.
The construction's 12-nearest-neighbor cutoff is a free parameter, and its sensitivity was not reported; a scan over k would test whether the triangle benefits are robust or tied to this specific graph density.

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.

Desk Editor's Note

read the letter

Quick take: QCformer is a genuine architectural contribution—a simplex Transformer running on quotient complexes of crystals, with triangle features standing in for three-body interactions—and it produces strong bandgap numbers on perovskite benchmarks. But the story about why the triangles help is weaker than the paper suggests, and the statistics are thin enough that several headline claims could be noise.

New and good: the quotient complex lifts a familiar idea (quotient graphs, used in MEGNet/ALIGNN/Matformer) to include 2-simplices, and the model updates node, edge, and triangle features with attention over neighbors and cofaces, using more layers at the node level then letting triangle updates flow down. JARVIS pretraining with transfer to HOIP is sensible and appears to pay off. Code is public. The authors are also straight about some limitations—they admit the HOIP2D MAE trails GDA-GBT and that the five-compound DFT check has functional-scatter problems.

Soft spots in order of size. First, no variance: every benchmark number is a single run, and on MP bandgap the edge over ComFormer is 0.191 vs 0.193 eV. The 2–11% ablation gains are exactly the kind of deltas that need repeated seeds. Second, the triangle ablation is only on MP and JARVIS, not on HOIP/HOIP2D, so the headline 30.59% HOIP improvement over MEGNet is attributed to triangles without a direct test. Third, the construction itself: Supplementary Figure S4 says the two 2-simplices in the ZnS quotient complex 'consist of two normal edges and one self-loop edge.' A 2-simplex with a self-loop is not a simplex; the homotopy-equivalent repair (eK, Lemma 4.3–Theorem 4.4) is proven, but the model consumes the glued object, including triangle features built from the self-loop length. That does not necessarily invalidate the empirical gains—it may still be a useful inductive bias—but the paper's 'higher-order interaction' language overclaims. Fourth, the five-compound validation is one outlier (452, error 0.68 eV) away from being a different story, and its 0.385 eV MAE is comparable to the 0.430 eV spread between GGA-PBE and r2SCAN; it doesn't settle the physical-meaningfulness question. Fifth, HOIP baselines are SchNet and MEGNet only—no Matformer or ComFormer—and pretraining is never ablated, so the source of the HOIP gain is underdetermined.

Worth noting: Theorem 4.5 is cited as Theorem S1.4 of the authors' own JCIM paper. That is an honest citation, but it means part of the 'mathematics' here is prior work, and the main text could say so more clearly.

Bottom line: the empirical claims are plausible, the representation is reusable, and the paper deserves a real referee. My recommendation: send it to review with requests for repeated-seed runs and significance tests, the triangle ablation on the perovskite datasets, a clarification of what exactly the model sees from the glued complex versus eK, and stronger HOIP baselines.

Referee Report

4 major / 4 minor

Summary. The paper introduces the Quotient Complex Transformer (QCformer), a topological deep learning model for predicting crystal and perovskite properties. A crystal is represented by taking the 12-nearest-neighbor graph of atoms in a unit cell, forming its clique complex, and gluing periodically equivalent atoms to obtain a quotient complex; vertex, edge, and triangle features are updated by a simplex-level Transformer with attention-based message passing. The model is pretrained on JARVIS and evaluated on Materials Project and JARVIS benchmarks, then fine-tuned on HOIP and HOIP2D bandgap datasets, where it reports state-of-the-art MAE values (0.0969 eV on HOIP, 0.0754 eV on HOIP2D), with ablations attributing part of the gain to triangle features. A theoretical section states that the quotient complex is homotopy equivalent to a true simplicial complex and describes its homology relative to the original crystal complex, but the proof is delegated to a prior paper by the same authors.

Significance. If the method's representational assumptions hold, this is a meaningful contribution: it is among the first applications of topological deep learning to materials property prediction, it combines periodic quotient structure with higher-order simplex features, and it reports consistent improvements over strong baselines on HOIP and HOIP2D benchmarks, with code and data availability stated. The reported benchmark numbers do support the central empirical claim that QCformer is competitive or better than ComFormer and Matformer on the tested inorganic tasks and substantially better on HOIP bandgap prediction. However, the significance is conditional on resolving whether the higher-order 'triangles' used by the model are well-defined geometric objects and whether the performance gain is reproducible beyond single runs; these points are load-bearing for the core novelty claim. No machine-checked proofs are supplied, and the central homology theorem is cited from the authors' prior work rather than proved in this manuscript.

major comments (4)

[§2.1.1 and Supplementary S2.1/S2.2 (Figs. S3–S4)] Supplementary Figure S4's caption states that after gluing periodic copies, 'the two 2-simplices consist of two normal edges and one self-loop edge.' A 2-simplex whose boundary contains a self-loop is not a simplex, so the quotient object used as the model input is not a simplicial complex; the paper's own repair (Lemma 4.3 and Theorem 4.4) constructs a homotopy-equivalent simplicial complex eK, but the model is not described as operating on eK. Since the triangle features defined in Supplementary S2.2 include products and squares of edge lengths, and self-loop edge lengths appear in those combinations, the claimed chemical meaning of the three-body features is not established. Please specify whether triangles with self-loop edges are kept, replace the model input by eK (or define the quotient complex as a cell complex), and add a sensitivity or physical validation that separates genuine three-body information from artifacts of degenerate triangles.
[§4.3.2 Tables 6–7; §2.3.1 Table 3] The triangle ablation (1D-QCformer vs QCformer) is only run on Materials Project and JARVIS, not on HOIP or HOIP2D. Consequently, the headline result of a 30.59% MAE improvement over MEGNet on the HOIP dataset is attributed in the text to higher-order triangle interactions without direct experimental support. Please run the ablation on the perovskite datasets, or explicitly state and justify that the gain transfers.
[§2.3.2 and Supplementary Table S2] The five-compound external validation does not resolve whether the model's predictions carry physical three-body information. The reported MAE of 0.3849 eV against GGA-PBE is comparable to the 0.4297 eV spread between GGA-PBE and r2SCAN, compound 452 alone contributes an absolute error of 0.6818 eV, and the text even notes that QCformer's prediction for compound 110 deviates strongly from both DFT functionals. With n=5 and no error bars, this experiment cannot discriminate the quotient-complex representation from a nontopological baseline.
[§4.2, Theorem 4.5; Supplementary S1.4] Theorem 4.5 is stated in the main text as a theorem justifying that the quotient complex captures periodic patterns, but its proof is not included in this manuscript; Supplementary Section S1.4 presents it as 'Theorem S1.4 (74)', where reference 74 is the authors' own prior paper. Please include a self-contained proof or a precise statement of the provenance, because this theorem is the theoretical basis for the claim that the quotient operation preserves the relevant topological information.

minor comments (4)

[Abstract; §2.3.1] The abstract says QCformer is pretrained on the Materials Project and JARVIS, but §2.3.1 states that pretraining for the perovskite tasks uses only the JARVIS Bandgap(OPT) task; please harmonize the wording.
[§2.1.2, Eq. (3)] In the attention coefficient formula, the query vector appears as [q_sigma, q_sigma], which looks like a typo for a concatenation involving q_sigma and a key- or message-related vector; please check.
[Supplementary S3.3] The text refers to 'Figure S3.3' while the figure is captioned 'Figure S5'; please fix the cross-reference.
[Tables 1–2] All results are reported without run-to-run variance; adding standard deviations over multiple seeds or folds would help the reader assess whether small deltas, such as 0.191 vs 0.193 eV on the MP bandgap task, are meaningful.

Circularity Check

1 steps flagged · score 2.0 of 10

One minor self-citation: the homology theorem behind the QC-periodicity claim is imported from the authors' own prior paper, while the reported prediction results are held-out and not circular.

self citation load bearing [Section 4.2 (Theorem 4.5); Supplementary S1.2 (Theorem S1.4); Reference [74]]
"Theorem S1.4 ( (74)) Let 𝐾,𝑉 = ˆ𝑘𝑗=1𝑉𝑗,K, and eK be defined as above. Then (a) 𝜃0 :𝐻0(K)→ 𝐻0( eK) is onto; (b) 𝜃1 :𝐻1(K)→ 𝐻1( eK) is one-to-one; (c) 𝜃𝑞 :𝐻𝑞(K)→ 𝐻𝑞( eK) is an isomorphism for𝑞 >1."

The main text presents Theorem 4.5 as the theoretical justification for the quotient-complex representation, stating that it 'demonstrates how the quotient complex captures the periodic patterns inherent in crystal structures.' However, the supplementary materials attribute the theorem to [74], the authors' own prior JCIM paper, and no proof is supplied in this manuscript. The theoretical claim that QCs encode periodicity is therefore imported from the authors' earlier work rather than derived here. This is a minor self-citation rather than a full circularity: the central empirical results are obtained from held-out evaluations on Materials Project, JARVIS, HOIP, and HOIP2D, and no fitted parameter is renamed as a prediction.

full rationale

The paper's central predictive claims are not circular: QCformer is trained and evaluated on external benchmarks with held-out splits for Materials Project, JARVIS, HOIP, and HOIP2D, and the triangle-feature ablation is a genuine comparison between 1D-QCformer and QCformer on held-out tasks. No fitted parameter is renamed as a prediction, and no prediction is defined in terms of the quantity it claims to predict. The only notable self-citation is the homology theorem (Theorem 4.5 / Theorem S1.4) used to justify that quotient complexes capture periodic patterns; it is cited from the authors' own prior paper rather than proved here. That theoretical support is not what generates the reported prediction accuracies, so it is a minor self-citation rather than a load-bearing circular derivation. Other concerns, such as the absence of error bars or the fact that the triangle ablation is not run on the HOIP datasets, are evidence-quality issues, not circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of hyperparameters, the kNN/clique construction, the transfer-learning assumption, and a homology theorem cited from the authors' prior work. The theorem is standard topology but is not proved in this paper; the kNN and triangle-feature choices are unvalidated physically; and the pretraining benefit is not ablated.

free parameters (5)

k-nearest neighbor count = 12
Determines graph connectivity and hence which triangles enter the quotient complex. Chosen by hand for all experiments; no sensitivity analysis is reported.
Simplex embedding dimension = 64
MLP maps vertex, edge, and triangle features to 64-dimensional embeddings. Set without reported tuning.
Message-passing layer counts = 5 edge-to-node layers + 2 triangle-to-edge-to-node layers
Architecture choice for hierarchical message passing; no ablation on layer counts is provided.
RBF center and scale parameters = Edges: c in [-4,0] step 1/64, sigma in {0.01,0.1,1}; triangles: c in [0,5] step 1/8, sigma in {0.01,0.1,1}
Feature expansion hyperparameters for edge and triangle embeddings; chosen manually and described in Section S2.2.
Learning rates and loss functions per dataset = lr 0.0006 with MSE for MP and JARVIS; lr 0.005 with MAE for HOIP and HOIP2D
Per-dataset training hyperparameters listed in Table S1; no sensitivity analysis.

assumptions (4)

standard math The quotient complex K obtained by gluing periodically equivalent atoms is homotopy equivalent to the simplicial complex eK built by coning off equivalence classes (Theorem 4.4).
Invoked in Section 4.2 to claim the quotient complex preserves homology. Cited to the authors' prior work (ref 74) and standard topology texts rather than proved in this paper.
standard math The homology of the quotient operation behaves as theta_0 onto, theta_1 injective, and theta_q isomorphic for q > 1 (Theorem 4.5).
Used to argue that quotient complexes capture periodic patterns while preserving higher-order topological information. Cited as Theorem S1.4 from ref 74, the authors' own prior paper.
domain assumption A 12-nearest-neighbor graph built from atomic coordinates in the unit cell, and its clique complex, faithfully represents the chemically relevant pairwise and three-body interactions for bandgap prediction.
Underpins all input structures. There is no physical validation that k=12 and clique triangles capture the relevant many-body interactions, and degenerate self-loop triangles are treated as valid 2-simplices (Supplementary Figure S4).
domain assumption Pretrained embeddings from JARVIS bandgap(OPT) transfer to HOIP and HOIP2D bandgap prediction.
The paper pretrains on 55,722 inorganic crystals and fine-tunes on perovskites, but no ablation is provided showing that pretraining improves over training from scratch (Section 2.3.1).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Quotient Complex Transformer (QCformer) for Perovskite Data Analysis." pith.science (2026). https://pith.science/paper/TEYLAKMC

@misc{pith2026250509174,
  author       = {Pith},
  title        = {Pith review of: Quotient Complex Transformer (QCformer) for Perovskite Data Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TEYLAKMC}},
  note         = {Machine review of arXiv:2505.09174}
}

read the original abstract

The discovery of novel functional materials is crucial in addressing the challenges of sustainable energy generation and climate change. Hybrid organic-inorganic perovskites (HOIPs) have gained attention for their exceptional optoelectronic properties in photovoltaics. Recently, geometric deep learning, particularly graph neural networks (GNNs), has shown strong potential in predicting material properties and guiding material design. However, traditional GNNs often struggle to capture the periodic structures and higher-order interactions prevalent in such systems. To address these limitations, we propose a novel representation based on quotient complexes (QCs) and introduce the Quotient Complex Transformer (QCformer) for material property prediction. A material structure is modeled as a quotient complex, which encodes both pairwise and many-body interactions via simplices of varying dimensions and captures material periodicity through a quotient operation. Our model leverages higher-order features defined on simplices and processes them using a simplex-based Transformer module. We pretrain QCformer on benchmark datasets such as the Materials Project and JARVIS, and fine-tune it on HOIP datasets. The results show that QCformer outperforms state-of-the-art models in HOIP property prediction, demonstrating its effectiveness. The quotient complex representation and QCformer model together contribute a powerful new tool for predictive modeling of perovskite materials.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 63 canonical work pages

[1]

Add all elements of 𝐴 to𝑆

work page
[2]

We have the inclusion map𝑖 : 𝐴→𝑆, and we denote the adjunction spaceK∪ 𝑖𝑆 by eK

For every pair of elements (𝑎1,𝑎 2) in 𝐴, if they satisfy 𝑎1∼𝑝𝑒𝑟𝑖𝑜𝑑 𝑎2, add a new vertex 𝑎12 and two edges(𝑎1,𝑎 12) and(𝑎2,𝑎 12) to𝑆. We have the inclusion map𝑖 : 𝐴→𝑆, and we denote the adjunction spaceK∪ 𝑖𝑆 by eK. It is noteworthy that eK remains a simplicial complex. Subsequently, we proceed to demonstrate that there exists a homotopy equivalence betwee...

work page
[3]

Pilania, C

G. Pilania, C. C. Wang, X. Jiang, S. Rajasekaran, R. Ramprasad, Accelerating materials property predictions using machine learning. Scientific reports 3, 2810 (2013)

work page 2013
[4]

Science, T

N. Science, T. C. (US),Materials genome initiative for global competitiveness(Executive Office of the President, National Science and Technology Council) (2011)

work page 2011
[5]

Jain, et al., The materials project: A materials genome approach to accelerating materials innovation, APL Mater

A. Jain, et al., The materials project: A materials genome approach to accelerating materials innovation, APL Mater. APL Materials (2013)

work page 2013
[6]

Choudhary, et al

K. Choudhary, et al. , The joint automated repository for various integrated simulations (JARVIS) for data-driven materials design.npj computational materials 6 (1), 173 (2020)

work page 2020
[7]

Himanen, et al., DScribe: Library of descriptors for machine learning in materials science

L. Himanen, et al., DScribe: Library of descriptors for machine learning in materials science. Computer Physics Communications 247, 106949 (2020)

work page 2020
[8]

M. Rupp, A. Tkatchenko, K.-R. M¨ uller, O. A. Von Lilienfeld, Fast and accurate modeling of molecular atomization energies with machine learning.Physical review letters108 (5), 058301 (2012)

work page 2012

Show all 79 references

[9]

H. Huo, M. Rupp, Unified representation of molecules and crystals for machine learning.arXiv preprint arXiv:1704.06439 (2017)

2017 arXiv
[10]

A. P. Bart ´ok, R. Kondor, G. Cs´anyi, On representing chemical environments. Physical Review B 87 (18), 184115 (2013)

2013
[11]

Behler, Atom-centered symmetry functions for constructing high-dimensional neural network potentials

J. Behler, Atom-centered symmetry functions for constructing high-dimensional neural network potentials. The Journal of chemical physics 134 (7) (2011)

2011
[12]

K. Atz, F. Grisoni, G. Schneider, Geometric deep learning on molecular representations.Nature Machine Intelligence 3 (12), 1023–1032 (2021)

2021
[13]

M. M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, P. Vandergheynst, Geometric deep learning: going beyond euclidean data. IEEE Signal Processing Magazine 34 (4), 18–42 (2017). 21

2017
[14]

M. M. Bronstein, J. Bruna, T. Cohen, P. Veli ˇckovi´c, Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478 (2021)

2021 arXiv
[15]

Masci, D

J. Masci, D. Boscaini, M. Bronstein, P. Vandergheynst, Geodesic convolutional neural networks on riemannian manifolds, in Proceedings of the IEEE international conference on computer vision workshops (2015), pp. 37–45

2015
[16]

Schmidt, et al., Predicting the thermodynamic stability of solids combining density functional theory and machine learning

J. Schmidt, et al., Predicting the thermodynamic stability of solids combining density functional theory and machine learning. Chemistry of Materials 29 (12), 5090–5103 (2017)

2017
[17]

C. Chen, W. Ye, Y. Zuo, C. Zheng, S. P. Ong, Graph networks as a universal machine learning framework for molecules and crystals.Chemistry of Materials 31 (9), 3564–3572 (2019)

2019
[18]

Choudhary, B

K. Choudhary, B. DeCost, Atomistic line graph neural network for improved materials property predictions. npj Computational Materials 7 (1), 185 (2021)

2021
[19]

Batzner, et al

S. Batzner, et al. , E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications 13 (1), 2453 (2022)

2022
[20]

T. Xie, J. C. Grossman, Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Physical review letters120 (14), 145301 (2018)

2018
[21]

C. W. Park, C. Wolverton, Developing an improved crystal graph convolutional neural network framework for accelerated materials discovery.Physical Review Materials4 (6), 063801 (2020)

2020
[22]

Louis, et al., Graph convolutional neural networks with global attention for improved materials property prediction

S.-Y. Louis, et al., Graph convolutional neural networks with global attention for improved materials property prediction. Physical Chemistry Chemical Physics 22 (32), 18141–18148 (2020)

2020
[23]

Schmidt, L

J. Schmidt, L. Pettersson, C. Verdozzi, S. Botti, M. A. Marques, Crystal graph attention networks for the prediction of stable materials. Science advances 7 (49), eabi7948 (2021)

2021
[24]

A. Y.-T. Wang, S. K. Kauwe, R. J. Murdock, T. D. Sparks, Compositionally restricted attention- based network for materials property predictions.Npj Computational Materials7 (1), 77 (2021)

2021
[25]

V. Fung, J. Zhang, E. Juarez, B. G. Sumpter, Benchmarking graph neural networks for materials chemistry. npj Computational Materials 7 (1), 84 (2021). 22

2021
[26]

K. Yan, Y. Liu, Y. Lin, S. Ji, Periodic graph transformers for crystal material property prediction. Advances in Neural Information Processing Systems35, 15066–15080 (2022)

2022
[27]

K. Yan, C. Fu, X. Qian, X. Qian, S. Ji, Complete and Efficient Graph Transformers for Crystal Material Property Prediction. arXiv preprint arXiv:2403.11857 (2024)

2024 arXiv
[28]

G. S. Na, Substructure interaction graph network with node augmentation for hybrid chemical systems of heterogeneous substructures.Computational Materials Science216, 111835 (2023)

2023
[29]

Deng, et al., CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling

B. Deng, et al., CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence5 (9), 1031–1041 (2023)

2023
[30]

Gruver, et al., Fine-Tuned Language Models Generate Stable Inorganic Materials as Text

N. Gruver, et al., Fine-Tuned Language Models Generate Stable Inorganic Materials as Text. arXiv preprint arXiv:2402.04379 (2024)

2024 arXiv
[31]

O. Kaba, S. Ravanbakhsh, Equivariant networks for crystal structures. Advances in Neural Information Processing Systems 35, 4150–4164 (2022)

2022
[32]

C. Chen, S. P. Ong, A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science 2 (11), 718–728 (2022)

2022
[33]

A. A. Duval, et al., Faenet: Frame averaging equivariant gnn for materials modeling, in Inter- national Conference on Machine Learning (PMLR) (2023), pp. 9013–9033

2023
[34]

Duval, et al., A Hitchhiker’s Guide to Geometric GNNs for 3D Atomic Systems

A. Duval, et al., A Hitchhiker’s Guide to Geometric GNNs for 3D Atomic Systems. arXiv preprint arXiv:2312.07511 (2023)

2023 arXiv
[35]

Fuchs, D

F. Fuchs, D. Worrall, V. Fischer, M. Welling, Se (3)-transformers: 3d roto-translation equiv- ariant attention networks. Advances in neural information processing systems 33, 1970–1981 (2020)

2020
[36]

Keriven, G

N. Keriven, G. Peyr ´e, Universal invariant and equivariant graph neural networks. Advances in Neural Information Processing Systems32 (2019)

2019
[37]

Wells, The geometrical basis of crystal chemistry

A. Wells, The geometrical basis of crystal chemistry. Part 1. Acta Crystallographica 7 (8-9), 535–544 (1954). 23

1954
[38]

A. F. Wells, Three-dimensional nets and polyhedra. (No Title) (1977)

1977
[39]

Iwamoto, Inclusion Compounds, ed

T. Iwamoto, Inclusion Compounds, ed. JL Atwood, JED Davies and DD MacNicol (1991)

1991
[40]

Iwamoto, S.-i

T. Iwamoto, S.-i. Nishikiori, T. Kitazawa, H. Yuge, Mineralomimetic chemistry as a modern aspect of co-ordination chemistry. Journal of the Chemical Society, Dalton Transactions 22, 4127–4136 (1997)

1997
[41]

Furukawa, J

H. Furukawa, J. Kim, N. W. Ockwig, M. O’Keeffe, O. M. Yaghi, Control of vertex geometry, structure dimensionality, functionality, and pore metrics in the reticular synthesis of crystalline metal- organic frameworks and Polyhedra.Journal of the American Chemical Society130 (35)...

2008
[42]

O’Keeffe, O

M. O’Keeffe, O. M. Yaghi, Deconstructing the crystal structures of metal–organic frameworks and related materials into their underlying nets. Chemical reviews 112 (2), 675–702 (2012)

2012
[43]

O. D. Friedrichs, M. O Keeffe, O. M. Yaghi, Three-periodic nets and tilings: regular and quasiregular nets. Acta Crystallographica Section A Foundations of Crystallography 59 (1), 22–27 (2003)

2003
[44]

Delgado-Friedrichs, M

O. Delgado-Friedrichs, M. O’Keeffe, Identification of and symmetry computation for crystal nets. Acta Crystallographica Section A: Foundations of Crystallography 59 (4), 351–360 (2003)

2003
[45]

O’Keeffe, M

M. O’Keeffe, M. A. Peskov, S. J. Ramsden, O. M. Yaghi, The reticular chemistry structure resource (RCSR) database of, and symbols for, crystal nets. Accounts of chemical research 41 (12), 1782–1789 (2008)

2008
[46]

S. J. Chung, T. Hahn, W. Klee, Nomenclature and generation of three-periodic nets: the vector method. Acta Crystallographica Section A: Foundations of Crystallography 40 (1), 42–50 (1984)

1984
[47]

Bader, W

M. Bader, W. Klee, G. Thimm, The 3-regular nets with four and six vertices per unit cell. Zeitschrift f¨ur Kristallographie-Crystalline Materials 212 (8), 553–558 (1997). 24

1997
[48]

Eon, Geometrical relationships between nets mapped on isomorphic quotient graphs: examples

J.-G. Eon, Geometrical relationships between nets mapped on isomorphic quotient graphs: examples. Journal of Solid State Chemistry 138 (1), 55–65 (1998)

1998
[49]

Klee, Crystallographic nets and their quotient graphs

W. Klee, Crystallographic nets and their quotient graphs. Crystal Research and Technology: Journal of Experimental and Industrial Crystallography 39 (11), 959–968 (2004)

2004
[50]

Sunada, Topological crystallography: with a view towards discrete geometric analysis, vol

T. Sunada, Topological crystallography: with a view towards discrete geometric analysis, vol. 6 (Springer Science & Business Media) (2012)

2012
[51]

Sunada, Lecture on topological crystallography

T. Sunada, Lecture on topological crystallography. Japanese Journal of Mathematics 7, 1–39 (2012)

2012
[52]

Eon, Topological features in crystal structures: a quotient graph assisted analysis of underlying nets and their embeddings

J.-G. Eon, Topological features in crystal structures: a quotient graph assisted analysis of underlying nets and their embeddings. Acta Crystallographica Section A: Foundations and Advances 72 (3), 268–293 (2016)

2016
[53]

Eon, From symmetry-labeled quotient graphs of crystal nets to coordination sequences: Algebraic tools for a combinatorial analysis of crystal structures

J.-G. Eon, From symmetry-labeled quotient graphs of crystal nets to coordination sequences: Algebraic tools for a combinatorial analysis of crystal structures. Structural Chemistry 23 (4), 987–996 (2012)

2012
[54]

Papillon, S

M. Papillon, S. Sanborn, M. Hajij, N. Miolane, Architectures of topological deep learning: A survey on topological neural networks. arXiv preprint arXiv:2304.10031 (2023)

2023 arXiv
[56]

S. Ebli, M. Defferrard, G. Spreemann, Simplicial neural networks. arXiv preprint arXiv:2010.03633 (2020)

2020 arXiv
[57]

E.-S. Kim, W. Y. Kang, K.-W. On, Y.-J. Heo, B.-T. Zhang, Hypergraph attention networks for multimodal learning, in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (2020), pp. 14581–14590

2020
[58]

Hajij, K

M. Hajij, K. Istvan, G. Zamzmi, Cell complex neural networks. arXiv preprint arXiv:2010.00743 (2020). 25

2020 arXiv
[59]

Bodnar, et al., Weisfeiler and lehman go topological: Message passing simplicial networks, in International Conference on Machine Learning (PMLR) (2021), pp

C. Bodnar, et al., Weisfeiler and lehman go topological: Message passing simplicial networks, in International Conference on Machine Learning (PMLR) (2021), pp. 1026–1037

2021
[60]

Y. Chen, Y. R. Gel, H. V. Poor, BScNets: Block simplicial complex neural networks, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36 (2022), pp. 6333–6341

2022
[61]

C. W. J. Goh, C. Bodnar, P. Lio, Simplicial attention networks.arXiv preprint arXiv:2204.09455 (2022)

2022 arXiv
[62]

S. H. Lee, F. Ji, W. P. Tay, SGAT: Simplicial graph attention network. arXiv preprint arXiv:2207.11761 (2022)

2022 arXiv
[63]

Hajij, et al., Higher-order attention networks

M. Hajij, et al., Higher-order attention networks. arXiv preprint arXiv:2206.00606 (2022)

2022 arXiv
[64]

Momma, F

K. Momma, F. Izumi, VESTA 3 for three-dimensional visualization of crystal, volumetric and morphology data. Journal of applied crystallography 44 (6), 1272–1276 (2011)

2011
[65]

Jain, et al., Commentary: The Materials Project: A materials genome approach to acceler- ating materials innovation

A. Jain, et al., Commentary: The Materials Project: A materials genome approach to acceler- ating materials innovation. Apl Materials 1 (1), 011002 (2013)

2013
[66]

Choudhary, B

K. Choudhary, B. DeCost, F. Tavazza, Machine learning with force-field-inspired descriptors for materials: Fast screening and mapping energy landscape. Phys. Rev. Mater. 2, 083801 (2018), doi:10.1103/PhysRevMaterials.2.083801, https://link.aps.org/doi/10.1103/ PhysRevMaterials...

2018 doi
[67]

C. Kim, T. D. Huan, S. Krishnan, R. Ramprasad, Data Descriptor: A hybrid organic-inorganic perovskite dataset (2017)

2017
[68]

E. I. Marchenko, et al., Database of two-dimensional hybrid perovskite materials: open-access collection of crystal structures, band gaps, and atomic partial charges predicted by machine learning. Chemistry of materials 32 (17), 7383–7388 (2020)

2020
[69]

D. V. Anand, Q. Xu, J. Wee, K. Xia, T. C. Sum, Topological feature engineering for machine learning based halide perovskite materials design. npj Computational Materials 8 (1), 203 (2022). 26

2022
[70]

C.-S. Hu, R. Mayengbam, M.-C. Wu, K. Xia, T. C. Sum, Geometric data analysis-based machine learning for two-dimensional perovskite design.Communications Materials 5 (1), 106 (2024)

2024
[71]

F. Mayr, A. Gagliardi, Global property prediction: a benchmark study on open-source, perovskite-like datasets. ACS omega 6 (19), 12722–12732 (2021)

2021
[72]

J. R. Munkres, Elements of algebraic topology (CRC Press) (2018)

2018
[73]

L¨ utgehetmann, D

D. L¨ utgehetmann, D. Govc, J. P. Smith, R. Levi, Computing persistent homology of directed flag complexes. Algorithms 13 (1), 19 (2020)

2020
[74]

J. W. Vick, Homology theory: an introduction to algebraic topology, vol. 145 (Springer Science & Business Media) (2012)

2012
[75]

Brown, Topology and Groupoids(www.groupoids.org) (2006), https://books.google

R. Brown, Topology and Groupoids(www.groupoids.org) (2006), https://books.google. com.sg/books?id=U9mMQAAACAAJ

2006
[76]

C.-S. Hu, R. Mayengbam, K. Xia, T. C. Sum, Quotient Complex (QC)-Based Machine Learning for 2D Hybrid Perovskite Design. Journal of Chemical Information and Modeling (2025)

2025
[77]

Paszke, et al., Pytorch: An imperative style, high-performance deep learning library

A. Paszke, et al., Pytorch: An imperative style, high-performance deep learning library. Ad- vances in neural information processing systems32 (2019)

2019
[78]

M. Fey, J. E. Lenssen, Fast graph representation learning with PyTorch Geometric. arXiv preprint arXiv:1903.02428 (2019)

2019 arXiv
[79]

Loshchilov, F

I. Loshchilov, F. Hutter, Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

2017 arXiv
[80]

L. N. Smith, N. Topin, Super-convergence: Very fast training of neural networks using large learning rates, in Artificial intelligence and machine learning for multi-domain operations applications (SPIE), vol. 11006 (2019), pp. 369–386. 27 Acknowledgments The authors thank Xia...

2019

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Add all elements of 𝐴 to𝑆

work page

[2] [2]

We have the inclusion map𝑖 : 𝐴→𝑆, and we denote the adjunction spaceK∪ 𝑖𝑆 by eK

For every pair of elements (𝑎1,𝑎 2) in 𝐴, if they satisfy 𝑎1∼𝑝𝑒𝑟𝑖𝑜𝑑 𝑎2, add a new vertex 𝑎12 and two edges(𝑎1,𝑎 12) and(𝑎2,𝑎 12) to𝑆. We have the inclusion map𝑖 : 𝐴→𝑆, and we denote the adjunction spaceK∪ 𝑖𝑆 by eK. It is noteworthy that eK remains a simplicial complex. Subsequently, we proceed to demonstrate that there exists a homotopy equivalence betwee...

work page

[3] [3]

Pilania, C

G. Pilania, C. C. Wang, X. Jiang, S. Rajasekaran, R. Ramprasad, Accelerating materials property predictions using machine learning. Scientific reports 3, 2810 (2013)

work page 2013

[4] [4]

Science, T

N. Science, T. C. (US),Materials genome initiative for global competitiveness(Executive Office of the President, National Science and Technology Council) (2011)

work page 2011

[5] [5]

Jain, et al., The materials project: A materials genome approach to accelerating materials innovation, APL Mater

A. Jain, et al., The materials project: A materials genome approach to accelerating materials innovation, APL Mater. APL Materials (2013)

work page 2013

[6] [6]

Choudhary, et al

K. Choudhary, et al. , The joint automated repository for various integrated simulations (JARVIS) for data-driven materials design.npj computational materials 6 (1), 173 (2020)

work page 2020

[7] [7]

Himanen, et al., DScribe: Library of descriptors for machine learning in materials science

L. Himanen, et al., DScribe: Library of descriptors for machine learning in materials science. Computer Physics Communications 247, 106949 (2020)

work page 2020

[8] [8]

M. Rupp, A. Tkatchenko, K.-R. M¨ uller, O. A. Von Lilienfeld, Fast and accurate modeling of molecular atomization energies with machine learning.Physical review letters108 (5), 058301 (2012)

work page 2012