Pith. sign in

REVIEW 6 major objections 4 minor 36 references

Crystal models get a new layer: coordination polyhedra as learnable graph nodes, cutting band-gap error to 0.292 eV on Materials Project data.

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

2026-08-01 22:17 UTC pith:NDK363YY

load-bearing objection The polyhedral-graph idea is worth a second look, but the paper's central mechanistic claim is untested: no ablations, single-seed results, and a contaminated band-gap target leave the 0.012 eV gain over ALIGNN unexplained. the 6 major comments →

arxiv 2607.24818 v1 pith:NDK363YY submitted 2026-07-17 cs.IR cond-mat.mtrl-scics.AIcs.CE

Dual-Level Atomic and Coordination Geometry Learning for Crystal Property Prediction Using Graph Neural Networks

Sanjay Chakraborty This is my paper
classification cs.IR cond-mat.mtrl-scics.AIcs.CE
keywords crystal property predictiongraph neural networkscoordination polyhedraVoronoi tessellationband gap predictionformation energymulti-scale representationmaterials informatics
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.

This paper argues that graph neural networks for crystal property prediction miss a physically central object: the coordination polyhedron, the atomic cage formed around each atom. The authors propose CPGN, a network that builds three coupled graphs — atoms, bonds, and polyhedra — and lets them exchange information through interleaved message passing and cross-attention. They claim this explicit geometric and topological prior improves predictions, particularly for the electronic band gap, where they report a Materials Project MAE of 0.292 eV, the best among the models they compare. If right, the paper establishes that polyhedral geometry and sharing topology are a useful inductive bias, not just decorative features.

Core claim

CPGN constructs a coordination polyhedron graph alongside the usual atom and line graphs. Each Voronoi-derived coordination shell becomes a node described by seven geometric descriptors (coordination number, bond-length distortion, angle variance, volume proxy, mean bond length, electronegativity mismatch, and coordination type), and edges between polyhedra encode whether they share a corner, edge, or face. The three graphs are updated in interleaved layers with bidirectional cross-attention between atom and polyhedron embeddings. The paper reports a formation-energy MAE of 0.060 eV/atom and the best band-gap MAE of 0.292 eV on the Materials Project, with competitive or better results on JAR

What carries the argument

The coordination polyhedron graph G_P: a coarse-grained graph where nodes are Voronoi tessellation shells (using Pymatgen's VoronoiNN within an 8 Å cutoff) and edges carry a scalar connectivity type — 0.0 for corner-sharing, 0.5 for edge-sharing, 1.0 for face-sharing. This graph injects explicitly encoded local geometry (distortion, angle variance, volume proxy, electronegativity mismatch) and inter-polyhedral topology into the learned representation, which the paper argues is what standard atom-only message passing must rediscover implicitly.

Load-bearing premise

The paper assumes that the Voronoi shells within an 8 Å cutoff are chemically meaningful coordination polyhedra, even though the mean coordination number at that cutoff (14.2) is well above typical chemical coordination numbers.

What would settle it

A reader could run the published CPGN code on the same Materials Project split but with a stricter Voronoi cutoff (e.g., 5 Å or a low-weight Voronoi factor) and compare band-gap MAE. If the polyhedral graph at a physically accurate coordination shell performs no better than the atom+line graph alone, the paper's central claim — that explicit polyhedral geometry drives the gain — is falsified.

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

If this is right

  • If CPGN's band-gap result holds, explicit polyhedral geometry becomes a validated alternative to deeper or higher-order message passing for geometry-sensitive electronic properties.
  • The three-graph architecture offers a template for injecting crystallographic prior knowledge into GNNs without abandoning atomic resolution.
  • The reported stability classifier (accuracy 0.988, F1 0.993) shows that sharing a single latent fingerprint across regression and classification heads can preserve physical consistency.
  • The parameter count (~4.85M on Materials Project) stays comparable to ALIGNN while adding a third graph stream, suggesting the polyhedral prior is relatively cheap to add.
  • On QM9, the model's HOMO/LUMO results indicate the same polyhedral machinery transfers to molecular orbital prediction, though with less gain over baselines than on crystalline band gaps.

Where Pith is reading between the lines

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

  • A natural testable extension is to ablate each of the seven polyhedron descriptors and each sharing-type edge weight, to see which geometric features actually drive the band-gap improvement — the paper does not report such ablations.
  • The 8 Å cutoff that yields a mean of 14.2 neighbours per atom may be too generous to represent true coordination polyhedra (typically 4–9 neighbours), so part of the reported gain could come from added model capacity rather than from chemically accurate polyhedra; a stricter cutoff or a learnable neighbor selection would settle this.
  • Because the polyhedron graph is built on Voronoi shells, its benefit likely depends on the quality of the input relaxed structures; applying CPGN to unrelaxed or disordered structures, which the paper lists as future work, would reveal how brittle the geometric prior is.
  • The cross-attention module is the largest parameter component (32.6%), so the paper's claim that gains come from polyhedral geometry rather than attention capacity could be probed by replacing cross-attention with a simple concatenation baseline.

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 / 4 minor

Summary. The manuscript proposes the Coordination Polyhedron Graph Network (CPGN), a multi-scale GNN that augments an atom graph and a line graph with a third graph whose nodes are Voronoi-derived coordination polyhedra and whose edges encode corner/edge/face sharing. The model alternates message passing on the three graphs, couples atom and polyhedron streams with scalar-gated cross-attention, and uses a shared latent vector for formation energy, band gap, and stability prediction. Results are reported on Materials Project 2018.6.1, JARVIS-DFT, and QM9; the headline numbers are a formation-energy MAE of 0.060 eV/atom and a band-gap MAE of 0.292 eV on Materials Project, with the band-gap result claimed as state of the art among the compared methods.

Significance. If the reported gains were robustly attributable to the polyhedron stream, CPGN would provide a useful inductive bias for geometry-sensitive crystal properties, and the paper's public code and detailed graph statistics would be a helpful resource. However, the central causal claim—that coordination-polyhedron geometry improves accuracy—is not currently isolated by controlled experiments. The architecture contribution is clearly described, but the evidence would need ablation studies, multi-seed uncertainty quantification, and clarification of the label-proxy issues before the claim can be accepted.

major comments (6)
  1. [§3 Steps 2–5; Tables 3, 7, 8, 13] The central claim that 'explicitly incorporating coordination polyhedra significantly enhances crystal representation learning' is not tested by ablation. The only comparisons are between the complete CPGN and independently trained or published baselines, which differ in parameter count (Table 10: CPGN 4.85M vs ALIGNN 4.03M), training protocol, and in Table 13 include numbers the authors state they could not reproduce. No experiment removes G_P, the 7-D descriptors, or the cross-attention while holding the backbone, capacity, and training protocol fixed. The observed differences (e.g., 0.012 eV on band gap) could therefore be due to model capacity or optimization rather than the polyhedral inductive bias. A parameter-matched ablation with the polyhedron stream disabled is necessary. The architecture description is also internally inconsistent: §4 (end of Step 4) says total trainable para
  2. [§4.1–4.3; all result tables] Every experiment uses a single random seed (SEED=42) and no standard deviations or seed ranges are reported. Since several headline improvements are small (band gap 0.292 vs ALIGNN 0.304; QM9 HOMO-LUMO gap 0.038 equal to ALIGNN; Materials Project formation energy tied with SchNet at 0.060), the reported ranking is not statistically supported. Multi-seed results with mean ± std should be provided for the main tables.
  3. [§3 Step 7; Table 3] The band-gap target includes the proxy max(0,−0.8E_f) for materials whose DFT band gap is unavailable. The fraction of test labels that are proxy values is not reported, and the proxy is a deterministic function of the formation-energy label used in the same multi-task objective (Eq. 10). For those samples the 'band-gap' prediction is effectively a recalibrated formation energy, so the 0.292 eV band-gap MAE may partly reflect learning of E_f rather than electronic structure. The authors should report the proxy fraction in train/test, give separate MAEs for true DFT gaps and proxy labels, and re-run the comparison with only true band-gap labels.
  4. [§4.3; Tables 3–4] The text reports validation MAE=0.0645 and test MAE=0.0666 eV/atom for formation energy, while the abstract, Table 3, and conclusions state 0.060 eV/atom. Similarly, §4.3 refers to 'CGCNN (23.85)' as a MAD:MAE ratio, whereas Table 4 lists CGCNN=11.7 and 23.85 is SSFormer's value in Table 13. These discrepancies concern the paper's central quantitative claims and must be resolved, with a clear statement of which split and checkpoint produce each reported number.
  5. [§3 Step 2; §5.2] The VoronoiNN shells at r_cut=8 Å yield a mean of 14.2 neighbors per atom, well above typical chemical coordination numbers, yet the descriptors normalize by 12 and cap CN at min(CN,12)/12. As written, the 'coordination polyhedra' may be effectively inflated Voronoi shells rather than chemically recognized coordination polyhedra. The authors should validate the construction (e.g., distribution of polyhedron node counts, comparison with known coordination numbers in reference structures, sensitivity to r_cut) to support the claim that the descriptors and sharing-type edges are physically interpretable.
  6. [§4.3.1; Tables 5–6] The stability label is 'obtained directly from the stability field... or approximated as 1 if E_f < 0, and 0 otherwise.' Since formation energy is the primary training target of the shared backbone, a stability classifier built on that backbone can achieve high accuracy by thresholding E_f. The reported accuracy/F1 (Table 5) and the comparison in Table 6 are therefore not a meaningful test of the polyhedral representation for stability, and the approximation should at least be reported and analyzed separately.
minor comments (4)
  1. [Figures and tables] Figure numbering is inconsistent: Figure 11 is used both for stability confusion matrices in §4.3.1 and for QM9 inference-time plots in §5.3. There is a floating 'Table??' reference near Table 6, and the Figure 2 caption contains a typo ('formaework').
  2. [§4.3 vs §4.5] The evaluation protocol for baselines is stated inconsistently: §4.3 says CGCNN, MEGNet, SchNet, and ALIGNN were reproduced, while §4.5 says QM9 baseline results were taken directly from the original papers. The authors should state, for each dataset, which baselines were rerun and under what split and hyperparameters.
  3. [§4.1 and §4.4] The MP split is described as a random shuffle with seed 42 (60,000/5,000/4,239), while some public benchmark comparisons use the standard MP split. If the published baseline numbers in Table 13 were obtained on a different split, the comparison is not controlled. Please state explicitly whether the same split was used for all baseline metrics.
  4. [§5.3; Table 12] The baseline inference times are 'estimated from published scaling characteristics' and are therefore only qualitative. The speed comparison should be labeled as approximate, or the baselines should be measured on the same hardware and code.

Circularity Check

2 steps flagged

Partial circularity: band-gap and stability evaluation targets are definitionally tied to the co-trained formation-energy target via proxies, so two reported prediction metrics partly reduce to formation-energy regression; the network/graph derivation and formation-energy benchmark are not circular.

specific steps
  1. self definitional [Section 3, Step 7 (Training), after Eq. (10); reported in Abstract and Tables 3/4]
    "The band-gap target is taken directly from the DFT-computed value in the dataset where available, with a proxy of max(0,−0.8E f ) applied only in its absence."

    For every entry lacking a DFT band gap, the target is the deterministic function max(0,−0.8E_f). Since E_f is the primary co-trained target (Eq. 10: L = MAE_Ef + 0.1·MAE_Eg), the band-gap head on those entries is scored against a rescaling of the formation-energy target. The advertised 'state-of-the-art' band-gap MAE of 0.292 eV is therefore partly a re-statement of formation-energy accuracy rather than independent evidence that coordination-polyhedron geometry improves band-gap prediction. The affected fraction is undisclosed, so the size of the reduction cannot be assessed from the paper.

  2. self definitional [Section 4.3.1 (Stability Classification), Eqs. (12)-(13) and Table 5]
    "The stability label is obtained directly from the stability field in the MP JSON file when available, or approximated as 1 if E f <0, and 0 otherwise. Because the stability prediction shares the same backbone as formation-energy regression ... which is why the classifier achieves aaccuracy of 0.989..."

    Where the MP stability field is unavailable, the label is defined as the sign of E_f, while E_f is the primary target of the same shared latent backbone. The stability head can therefore succeed by thresholding the formation-energy prediction; the reported 0.9882 accuracy and 0.9913 ROC-AUC are partly a check on the formation-energy regressor, not independent evidence that the coordination-polyhedron representation encodes thermodynamic stability. The paper's own 'this is why ... 0.989' sentence acknowledges the dependence, yet the number is presented as a CPGN stability result.

full rationale

The core derivation (CIF/POSCAR to atom/line/polyhedron graphs, message passing, cross-attention, and the formation-energy head) is self-contained: the 7-D polyhedron descriptors are computed from structures, not from targets, and no fitted parameter is renamed as a prediction. The formation-energy MAE of 0.060 eV/atom is anchored to the standard MP 2018.6.1 split with reproduced baselines, so the primary benchmark retains independent content. There is no load-bearing self-citation chain. The circularity is confined to the auxiliary prediction targets: for an undisclosed subset, Eg and stability are defined as functions of the co-trained Ef (max(0,−0.8Ef) and 1_{Ef<0}), making the reported band-gap SOTA and stability accuracy partly self-referential re-statements of formation-energy accuracy. Because the band-gap result is one of the paper's two headline findings, this is partial circularity (not full: the graph-construction and formation-energy claims are not reduced to their inputs). Data-quality issues such as the lack of an ablation, single-seed runs, and parameter-count differences are evidentiary weaknesses, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The architecture's load-bearing assumptions are the identification of Voronoi shells with coordination polyhedra, the sufficiency of the 7 hand-crafted descriptors, the surrogate band-gap labels derived from formation energy, and the representativeness of a single seed-42 split. All are domain assumptions or ad hoc choices that the paper does not test with ablations or repeated-seed statistics.

free parameters (7)
  • Voronoi/bond cutoff r_cut = 8.0 Å
    Chosen by hand (§3 Step 2); defines which neighbors enter the atom graph, line-graph triplets, and the Voronoi coordination shells that become polyhedron nodes. Mean coordination becomes 14.2, suggesting the cutoff inflates shells above traditional coordination numbers.
  • Auxiliary loss weight λ_aux = 0.1
    Fixed by hand (§4.1, Eq. 10); controls how much band-gap and stability signals shape the shared latent vector; the paper notes the choice is not optimized.
  • Polyhedron connectivity weights = corner 0.0, edge 0.5, face 1.0
    Hand-assigned scalar encodings of sharing type (§3 Step 3); no data-driven or physically calibrated basis is given.
  • Band-gap proxy coefficients = max(0, −0.8·E_f)
    Used as band-gap label where DFT gap is absent (§3 Step 7); this is a fitted linear map from formation energy to gap and contaminates the headline 0.292 eV MAE with an unknown fraction of surrogate labels.
  • Polyhedron descriptor normalizations = CN/12, angle variance/10000, d̄³/100, d̄/r_cut
    Hand-chosen scales (§3 Step 2); features (i) and (vii) are near-duplicates (CN/12 vs min(CN,12)/12), so the 7-D vector carries less independent information than advertised.
  • Training seed and hyperparameters = seed 42, LR 3e-4, weight decay 1e-5, L=4, d_h=256, patience 50
    Single fixed seed; no repeated runs or error bars, yet differences of 0.012–0.014 eV are claimed as improvements (§4.1, Tables 3-4).
  • Stability class weight = 3.0
    Positive-class weight for BCE (§4.3.1); hand-set, and the stability task is admitted to be nearly saturated by class imbalance (96.1% stable).
axioms (6)
  • domain assumption VoronoiNN shells within 8 Å equal chemically meaningful coordination polyhedra (Section 3, Step 2).
    All polyhedron nodes, descriptors, and sharing edges derive from this identification; Section 5.2's mean coordination of 14.2 suggests shells include weak or artificial contacts beyond the usual coordination polyhedron.
  • ad hoc to paper The seven hand-crafted geometric descriptors capture the physics relevant to formation energy, band gap, and stability (Section 3, Step 2).
    No feature-importance or ablation analysis is provided; normalization constants are arbitrary (CN/12, variance/10000, d̄³/100); two of seven features are near-duplicates of the coordination number.
  • domain assumption DFT labels in MP 2018.6.1, JARVIS-DFT, and QM9 are ground truth for training and evaluation (Section 4).
    Standard for the field; not independently verified here, and the MP band-gap field is partially replaced by a surrogate label.
  • domain assumption A random seed-42 split yields i.i.d. validation/test sets with no leakage (Section 4.1).
    Split is structure-based random; no composition- or cluster-based splitting is discussed; single realization.
  • domain assumption Line-graph edges between bonds sharing a destination atom capture bond-angle geometry (Section 3, Step 3).
    Borrowed from ALIGNN [5]; reasonable but inherited without re-derivation.
  • domain assumption A single training run at seed 42 is representative of model performance (Section 4.1, Table 2).
    No repeated-seed statistics; reported gains (0.012–0.014 eV) are likely within run-to-run variance for these models.
invented entities (3)
  • Coordination polyhedron graph G_P no independent evidence
    purpose: Nodes are Voronoi coordination shells; edges encode corner/edge/face-sharing connectivity; provides the coarse-grained geometric stream claimed to drive gains.
    The paper gives no ablation, sensitivity analysis, or external validation isolating G_P's contribution; its input features depend on the unvalidated Voronoi-shell assumption (Sec. 3, Step 3).
  • Scalar-gated bidirectional cross-attention between atom and polyhedron streams no independent evidence
    purpose: Couples atom- and polyhedron-level embeddings at each of 4 layers; the largest parameter block (32.6% of total).
    No ablation removes or varies this module; the paper itself notes the cost comes from architecture design rather than graph structure (Sec. 5.1), so its benefit is unmeasured.
  • 7-D polyhedron descriptor vector no independent evidence
    purpose: Encodes coordination number, distortion, angle variance, volume proxy, mean bond length, electronegativity mismatch, and coordination type as learnable geometric priors.
    No feature-importance, perturbation, or interpretability analysis is provided; two components are functions of the same coordination-number value.

pith-pipeline@v1.3.0-alltime-deepseek · 158 in / 20395 out tokens · 191619 ms · 2026-08-01T22:17:39.901740+00:00 · methodology

0 comments
read the original abstract

Accurate prediction of crystal properties remains a key challenge in computational materials science. While graph neural networks (GNNs) such as CGCNN, MEGNet, ALIGNN, and SchNet have shown strong performance, they primarily represent crystals at the atomic level and implicitly learn local chemical environments through message passing. However, many material properties are governed by coordination polyhedra, the fundamental structural units formed by atoms and their neighboring atoms. To address this limitation, we propose the Coordination Polyhedron Graph Network (CPGN), a multi-scale GNN that jointly learns atomic, bond, and coordination-polyhedron representations. CPGN constructs three coupled graphs: an atom graph encoding elemental and bonding information, a line graph capturing angular interactions, and a coordination polyhedron graph describing Voronoi-derived local environments through corner-, edge-, and face-sharing relationships. Physically meaningful geometric descriptors are incorporated for each polyhedron, while an interleaved message-passing mechanism with bidirectional cross-attention enables effective information exchange across structural levels. Extensive evaluations on the Materials Project, JARVIS-DFT, and QM9 benchmark datasets demonstrate that CPGN outperforms existing state-of-the-art GNN models. It achieves a formation-energy MAE of 0.060 eV/atom and a band-gap MAE of 0.292 eV on the Materials Project, while providing competitive multi-property prediction on JARVIS-DFT and superior HOMO prediction on QM9. The results highlight that explicit modeling of coordination polyhedra improves crystal representation learning and enables accurate, physically interpretable prediction of material properties.

Figures

Figures reproduced from arXiv: 2607.24818 by Sanjay Chakraborty.

Figure 1
Figure 1. Figure 1: Workflow of the proposed CPGN framework Step 1 — The input: a crystal blueprint The entire workflow begins with a crystal structure file, typically in CIF or POSCAR format, which serves as a precise digital blueprint of the material under investigation. This file encodes the identity of every atom in the repeating unit of the crystal, the three-dimensional fractional coordinates of each atomic site, the sh… view at source ↗
Figure 2
Figure 2. Figure 2: Performance plots of the proposed CPGN formaework on MPProject dataset [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance plots of SchNet on MPProject dataset [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance plots of ALIGNN on MPProject dataset [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance plots of MEGNet on MPProject dataset [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance plots of CPGN stability classification on MPProject dataset [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance plots of CPGN on JARVIS-DFT dataset on 37,711 materials [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance plots of CGCNN on JARVIS-DFT dataset on 37,711 materials [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance plots of ALIGNN on JARVIS-DFT dataset on 37,711 materials [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance plots of CPGN on QM9 dataset for HOMO and LUMO [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Performance plots of CPGN on QM9 dataset for HOMO and LUMO [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 2 linked inside Pith

  1. [1]

    Quantifying uncertainty in high- throughput density functional theory: A comparison of aflow, materials project, and oqmd.Physical Review Materials, 7(5):053805, 2023

    Vinay I Hegde, Christopher KH Borg, Zachary Del Rosario, Yoolhee Kim, Maxwell Hutchinson, Erin Antono, Julia Ling, Paul Saxe, James E Saal, and Bryce Meredig. Quantifying uncertainty in high- throughput density functional theory: A comparison of aflow, materials project, and oqmd.Physical Review Materials, 7(5):053805, 2023. 21

    2023

  2. [2]

    Recent advances and applications of deep learning methods in materials science.npj Computational Materials, 8(1):59, 2022

    Kamal Choudhary, Brian DeCost, Chi Chen, Anubhav Jain, Francesca Tavazza, Ryan Cohn, Cheol Woo Park, Alok Choudhary, Ankit Agrawal, Simon JL Billinge, et al. Recent advances and applications of deep learning methods in materials science.npj Computational Materials, 8(1):59, 2022

    2022

  3. [3]

    Graph convolutional neural networks with global attention for improved materials property prediction.Physical Chemistry Chemical Physics, 22(32):18141–18148, 2020

    Steph-Yves Louis, Yong Zhao, Alireza Nasiri, Xiran Wang, Yuqi Song, Fei Liu, and Jianjun Hu. Graph convolutional neural networks with global attention for improved materials property prediction.Physical Chemistry Chemical Physics, 22(32):18141–18148, 2020

    2020

  4. [4]

    Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Physical review letters, 120(14):145301, 2018

    Tian Xie and Jeffrey C Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Physical review letters, 120(14):145301, 2018

    2018

  5. [5]

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

    Kamal Choudhary and Brian DeCost. Atomistic line graph neural network for improved materials property predictions.npj Computational Materials, 7(1):185, 2021

    2021

  6. [6]

    Graph networks as a universal machine learning framework for molecules and crystals.Chemistry of Materials, 31(9):3564–3572, 2019

    Chi Chen, Weike Ye, Yunxing Zuo, Chen Zheng, and Shyue Ping Ong. Graph networks as a universal machine learning framework for molecules and crystals.Chemistry of Materials, 31(9):3564–3572, 2019

    2019

  7. [7]

    Periodic graph transformers for crystal material property prediction.Advances in Neural Information Processing Systems, 35:15066–15080, 2022

    Keqiang Yan, Yi Liu, Yuchao Lin, and Shuiwang Ji. Periodic graph transformers for crystal material property prediction.Advances in Neural Information Processing Systems, 35:15066–15080, 2022

    2022

  8. [8]

    Recinet: Reciprocal space-aware long-range modeling for crystalline property prediction.arXiv preprint arXiv:2502.02748, 2025

    Jianan Nie, Peiyao Xiao, Kaiyi Ji, and Peng Gao. Recinet: Reciprocal space-aware long-range modeling for crystalline property prediction.arXiv preprint arXiv:2502.02748, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Pscg-net: A multiscale crystal graph neural network for accelerated materials discovery.Journal of Chemical Information and Modeling, 65(20):10871–10884, 2025

    Guangyao Chen, Zhilong Wang, and Fengqi You. Pscg-net: A multiscale crystal graph neural network for accelerated materials discovery.Journal of Chemical Information and Modeling, 65(20):10871–10884, 2025

    2025

  10. [10]

    From polyhedra to crystals: a graph- theoretic framework for crystal structure generation.CrystEngComm, 2026

    Tomoyasu Yokoyama, Kazuhide Ichikawa, and Hisashi Naito. From polyhedra to crystals: a graph- theoretic framework for crystal structure generation.CrystEngComm, 2026

    2026

  11. [11]

    Local coordina- tion versus overall topology in crystal structures: deriving knowledge from crystallographic databases

    Alexander P Shevchenko, Igor A Blatov, Elena V Kitaeva, and Vladislav A Blatov. Local coordina- tion versus overall topology in crystal structures: deriving knowledge from crystallographic databases. Crystal Growth & Design, 17(2):774–785, 2017

    2017

  12. [12]

    Commentary: The materials project: A materials genome approach to accelerating materials innovation.APL Materials, 1:011002, 2013

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, et al. Commentary: The materials project: A materials genome approach to accelerating materials innovation.APL Materials, 1:011002, 2013

    2013

  13. [13]

    Schnet– a deep learning architecture for molecules and materials.The Journal of Chemical Physics, 148(24), 2018

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

    2018

  14. [14]

    Examining graph neural networks for crystal structures: Limitations and opportunities for capturing periodicity.Science Advances, 9:eadi3245, 2023

    Cong Fu, Handong Ye, Meng Wen, Yining Feng, Kerk Schwarz, Yuyang Sun, Mihir Bhatt, Tarak Bhatt, Subhajit Sanyal, Zhirong Deng, et al. Examining graph neural networks for crystal structures: Limitations and opportunities for capturing periodicity.Science Advances, 9:eadi3245, 2023

    2023

  15. [15]

    Garrity, Andrew C

    Kamal Choudhary, Kevin F. Garrity, Andrew C. E. Reid, Brian DeCost, Adam J. Biacchi, Angela R. Hight Walker, Zachary Starrett, Igor Buda, Igal Fleischer, James Park, et al. The joint automated repository for various integrated simulations (jarvis) for data-driven materials design.npj Computa- tional Materials, 6:173, 2020

    2020

  16. [16]

    Quantum chemistry structures and properties of 134 kilo molecules.Scientific data, 1(1):1–7, 2014

    Raghunathan Ramakrishnan, Pavlo O Dral, Matthias Rupp, and O Anatole Von Lilienfeld. Quantum chemistry structures and properties of 134 kilo molecules.Scientific data, 1(1):1–7, 2014

    2014

  17. [17]

    Fast and uncertainty-aware directional message passing for non-equilibrium molecules.NeurIPS 2020 Workshop on Machine Learning for Molecules, 2020

    Johannes Gasteiger, Florian Becker, and Stephan G¨ unnemann. Fast and uncertainty-aware directional message passing for non-equilibrium molecules.NeurIPS 2020 Workshop on Machine Learning for Molecules, 2020

    2020

  18. [18]

    Gemnet: Universal directional graph neural networks for molecules.Advances in Neural Information Processing Systems, 34:6790–6802, 2021

    Johannes Gasteiger, Florian Becker, and Stephan G¨ unnemann. Gemnet: Universal directional graph neural networks for molecules.Advances in Neural Information Processing Systems, 34:6790–6802, 2021

    2021

  19. [19]

    Spherical message passing for 3d molecular graphs.International Conference on Learning Representations (ICLR), 2022

    Yi Liu, Limei Wang, Meng Liu, Xuan Zhang, Bora Oztekin, and Shuiwang Ji. Spherical message passing for 3d molecular graphs.International Conference on Learning Representations (ICLR), 2022. 22

    2022

  20. [20]

    Heilman, Weiyi Gong, and Qimin Yan

    Alexander J. Heilman, Weiyi Gong, and Qimin Yan. Crystal hypergraph convolutional networks.arXiv preprint arXiv:2411.12616, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    A universal graph deep learning interatomic potential for the periodic table.Nature Computational Science, 2:718–728, 2022

    Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table.Nature Computational Science, 2:718–728, 2022

    2022

  22. [22]

    Bartel, and Gerbrand Ceder

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Gerbrand Ceder. Chgnet as a pretrained universal neural network potential for charge-informed atom- istic modelling.Nature Machine Intelligence, 5:1031–1041, 2023

    2023

  23. [23]

    Ilyes Batatia, D´ avid P´ eter Kov´ acs, Gregor N. C. Simm, Christoph Ortner, and G´ abor Cs´ anyi. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields.Advances in Neural Information Processing Systems, 35:11423–11436, 2022

    2022

  24. [24]

    High-throughput identification and characterization of two-dimensional materials using density functional theory.Scientific reports, 7(1):5179, 2017

    Kamal Choudhary, Irina Kalish, Ryan Beams, and Francesca Tavazza. High-throughput identification and characterization of two-dimensional materials using density functional theory.Scientific reports, 7(1):5179, 2017

    2017

  25. [25]

    Computational screening of high-performance optoelectronic materials using optb88vdw and tb-mbj formalisms.Scientific data, 5(1):180082, 2018

    Kamal Choudhary, Qin Zhang, Andrew CE Reid, Sugata Chowdhury, Nhan Van Nguyen, Zachary Trautt, Marcus W Newrock, Faical Yannick Congo, and Francesca Tavazza. Computational screening of high-performance optoelectronic materials using optb88vdw and tb-mbj formalisms.Scientific data, 5(1):180082, 2018

    2018

  26. [26]

    Elastic properties of bulk and low-dimensional materials using van der waals density functional.Physical review

    Kamal Choudhary, Gowoon Cheon, Evan Reed, and Francesca Tavazza. Elastic properties of bulk and low-dimensional materials using van der waals density functional.Physical review. B, 98(1):10–1103, 2018

    2018

  27. [27]

    Accelerated discovery of efficient solar cell materials using quantum and machine-learning methods

    Kamal Choudhary, Marnik Bercx, Jie Jiang, Ruth Pachter, Dirk Lamoen, and Francesca Tavazza. Accelerated discovery of efficient solar cell materials using quantum and machine-learning methods. Chemistry of materials, 31(15):5900–5908, 2019

    2019

  28. [28]

    Machine learning with force-field-inspired descriptors for materials: Fast screening and mapping energy landscape.Physical review materials, 2(8):083801, 2018

    Kamal Choudhary, Brian DeCost, and Francesca Tavazza. Machine learning with force-field-inspired descriptors for materials: Fast screening and mapping energy landscape.Physical review materials, 2(8):083801, 2018

    2018

  29. [29]

    High-throughput discovery of topologically non-trivial materials using spin-orbit spillage.Scientific reports, 9(1):8534, 2019

    Kamal Choudhary, Kevin F Garrity, and Francesca Tavazza. High-throughput discovery of topologically non-trivial materials using spin-orbit spillage.Scientific reports, 9(1):8534, 2019

    2019

  30. [30]

    The joint automated repository for various integrated simulations (jarvis) for data-driven materials design

    Kamal Choudhary, Kevin F Garrity, Andrew CE Reid, Brian DeCost, Adam J Biacchi, Angela R Hight Walker, Zachary Trautt, Jason Hattrick-Simpers, A Gilad Kusne, Andrea Centrone, et al. The joint automated repository for various integrated simulations (jarvis) for data-driven materials design. npj computational materials, 6(1):173, 2020

    2020

  31. [31]

    Prediction errors of molecular machine learning models lower than hybrid dft error.Journal of chemical theory and computation, 13(11):5255–5264, 2017

    Felix A Faber, Luke Hutchison, Bing Huang, Justin Gilmer, Samuel S Schoenholz, George E Dahl, Oriol Vinyals, Steven Kearnes, Patrick F Riley, and O Anatole Von Lilienfeld. Prediction errors of molecular machine learning models lower than hybrid dft error.Journal of chemical theory and computation, 13(11):5255–5264, 2017

    2017

  32. [32]

    Fast and uncertainty-aware directional message passing for non-equilibrium molecules.Machine Learning for Molecules Workshop at NeurIPS, 2020

    Johannes Gasteiger, Shankari Giri, Johannes T Margraf, and Stephan G¨ unnemann. Fast and uncertainty-aware directional message passing for non-equilibrium molecules.Machine Learning for Molecules Workshop at NeurIPS, 2020

    2020

  33. [33]

    Strategies for pre-training graph neural networks.International Conference on Learning Representa- tions (ICLR), 2020

    Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. Strategies for pre-training graph neural networks.International Conference on Learning Representa- tions (ICLR), 2020

    2020

  34. [34]

    Ssformer: Self-supervised transformer model for predicting the properties of crystalline materials.Computational Materials Science, 263:114447, 2026

    Jie Feng, Tao Sun, and Jianmei Yuan. Ssformer: Self-supervised transformer model for predicting the properties of crystalline materials.Computational Materials Science, 263:114447, 2026

    2026

  35. [35]

    Jiahao Kang, Liang Yang, Jingjie Zeng, Zhi Sun, Hongfei Lin, and Junpeng Li. Graph isomorphism attention network combined with pre-trained language models: a novel approach for crystal material property prediction.Neural Computing and Applications, 37(30):25061–25076, 2025

    2025

  36. [36]

    Benchmarking materi- als property prediction methods: the matbench test set and automatminer reference algorithm.npj Computational Materials, 6(1):138, 2020

    Alexander Dunn, Qi Wang, Alex Ganose, Daniel Dopp, and Anubhav Jain. Benchmarking materi- als property prediction methods: the matbench test set and automatminer reference algorithm.npj Computational Materials, 6(1):138, 2020. 23

    2020