Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Material Property Prediction with Element Attribute Knowledge Graphs and Multimodal Representation Learning

T0 review · 4 major / 5 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 This paper claims that adding element-property knowledge-graph embeddings to crystal-graph features lowers band-gap prediction error to 0.177 eV on the standard benchmark.

desk verdict A plausible incremental fusion of element knowledge graphs with crystal graphs, but the headline gain is not cleanly attributed and the ablation is too weak to support 'significantly improves.' read the letter →

arxiv 2411.08414 v1 pith:267J7QQX submitted 2024-11-13 cs.LG cond-mat.mtrl-scics.AI

Chao Huang , Chunyan Chen , Ling Shi , Chen Chen This is my paper · ORCID
classification cs.LGcond-mat.mtrl-scics.AI
keywords materialpropertypredictionknowledgegraphelementattributesmultimodalrepresentationlearningbandgapformationenergycrystalstructureneuralnetwork
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

Machine-learning models for crystalline materials usually encode atomic positions and bonding, but not the chemical identity of the elements themselves. This paper argues that element attributes, such as electronegativity, atomic radius, and ionization energy, carry signal that structure alone misses. To test that idea, ESNet builds an element knowledge graph, embeds it with a semantic embedding model, and merges the resulting element vector with the crystal-graph representation from a transformer encoder. On the standard benchmark, the fused model reports a band-gap mean absolute error of 0.177 eV, improving on the 0.193 eV of its structure-only backbone, while formation-energy error remains comparable. The conclusion the paper draws is that chemical prior knowledge and geometric structure are complementary for crystal property prediction.

What carries the argument

The central object is the element knowledge graph: triples of the form (attribute value, is attribute of, element), built from periodic-table data and discretized for continuous quantities. It is embedded with OWL2Vec, a semantic-embedding method that converts the ontology into a graph, runs random walks, collects structural, lexical, and combinatorial documents, and trains Word2Vec to produce element vectors. The crystal side uses iComFormer, a periodic graph transformer with SE(3)-invariant crystal graphs, using a redesigned 70-dimensional atom-feature vector that replaces CGCNN's 92-dimension encoding and corrects errors for hydrogen and heavy elements. The two representations are combined by computing the stoichiometric weighted sum of element vectors to form $H_e$ and then applying the linear fusion $H_f = \alpha H_e + \beta H_g$, followed by task-specific transformer layers.

What would settle it

Run ESNet and its no-knowledge-graph ablation many times with different random seeds on the same benchmark split; if the 0.177 eV versus 0.182 eV gap does not consistently exceed seed-level noise, the claim that element-graph fusion improves prediction is falsified. A sharper test is to replace the learned element embeddings with fixed random vectors of the same dimension: if band-gap MAE stays at 0.177 eV, the knowledge graph itself is not doing the work.

Watch

Extended reading notes

Core claim

The central claim is that the element knowledge graph is not merely a feature-engineering trick but a genuinely complementary modality. The element encoder represents each element by an embedding learned from a knowledge graph of element-attribute triples; ESNet combines the embeddings of a crystal's constituent elements in proportion to their stoichiometric counts to form $H_e$, and the crystal graph encoder produces $H_g$ from atomic positions and bond geometry. The fusion encoder forms $H_f = \alpha H_e + \beta H_g$ and passes the joint representation through transformer layers before predicting band gap or formation energy. With this design the paper reports the lowest band-gap MAE (0.177 eV) among the ten methods it compares, and its ablation removing the element KG encoder raises the MAE to 0.182 eV. The paper frames this as evidence that adding element-level chemical attributes significantly improves performance, particularly for band-gap prediction.

Load-bearing premise

The load-bearing premise is that the element-knowledge-graph vector carries chemical information that is complementary to the crystal-structure vector under the weighted sum $H_f = \alpha H_e + \beta H_g$; if the improvement actually comes from the redesigned 70-dimensional atom features or from hyperparameter choices, the multimodal-fusion claim would not hold.

Editorial extensions

If this is right

  • Band-gap prediction on the standard benchmark improves from 0.193 eV to 0.177 eV MAE when element-knowledge features are fused with the crystal graph.
  • Removing the element KG encoder increases band-gap MAE to 0.182 eV, indicating that the fusion component, not just the structural encoder, contributes to the reported accuracy.
  • The redesigned 70-dimensional atom features retain or improve on the 92-dimensional CGCNN encoding, so element attributes can be integrated without sacrificing structural information.
  • The same fused representation can be applied to other target properties: on formation energy the model matches existing benchmarks rather than exceeding them.
  • Because the element embeddings are obtained independently of the crystal encoder, the element-knowledge component can be reused with different structural backbones or downstream tasks.

Reading between the lines

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

  • Beyond the paper, if the element KG is genuinely complementary, the same element embeddings could be plugged into other crystal encoders, such as line-graph or message-passing models, and should transfer most of the reported gain.
  • The ablation improvement is 0.005 eV, a difference that is within run-to-run noise on many graph benchmarks; an independent multi-seed comparison would be needed to separate the KG signal from the redesigned atom features.
  • The stoichiometric weighted sum treats a material as a bag of element proportions; a site-aware fusion that attaches element embeddings to individual atoms could capture ordering and local composition effects the current method flattens.
  • If the gain is real, element-level chemical priors could be most valuable in low-data regimes, where structural graph training is data-hungry; the paper does not evaluate few-shot or extrapolation settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents ESNet, a multimodal framework for crystal property prediction that fuses element-level knowledge graph embeddings (derived from periodic-table triples via OWL2Vec/Word2Vec) with crystal graph features from iComFormer, using a Transformer-based fusion encoder. On the Materials Project benchmark, the method reports a band-gap MAE of 0.177 eV versus 0.193 eV for iComFormer, and a formation-energy MAE of 20.05 meV/atom versus 18.26 meV/atom. The central claim is that the joint representation of element attributes and crystal structure significantly improves performance, particularly for band-gap prediction.

Significance. If the reported band-gap improvement is real and attributable to the element knowledge graph branch, ESNet would be a useful demonstration that chemically grounded element-level priors can complement structural graph encoders. The paper also contributes a redesigned 70-dimensional atomic feature vector that appears to improve on the standard CGCNN 92-dimensional encoding, and the use of an unsupervised element knowledge graph constructed from independent periodic-table data avoids circularity. However, the evidence for the central multimodal-fusion claim is currently weak because the comparison with iComFormer is not controlled, the fusion equation is inconsistent with the prose, and the only ablation reports a 0.005 eV gap without uncertainty quantification.

major comments (4)
  1. [§3.3, Eq. (1)] Equation (1) states H_f = αH_e + βH_g, but the preceding text says the fusion encoder 'concatenate[s] the elemental property features H_e ... and the geometric features ... H_g to create a joint representation.' Concatenation would produce a vector of dimension d1+d2, not a weighted sum. Furthermore, α and β are never specified numerically or described as learned per-task coefficients; they are free parameters whose values are essential for reproducibility of the fused representation. Please clarify the actual operation, the dimensions involved, and how α and β are determined.
  2. [§4.2 and §3.2] The ablation in Table 4 removes only the element KG encoder, but it retains the redesigned 70-dimensional atom features introduced in §3.2. Meanwhile, the iComFormer baseline in Table 2 uses the original 92-dimensional CGCNN encoding. Consequently, the reported band-gap improvement of ESNet over iComFormer (0.177 vs 0.193 eV) conflates two distinct changes: the 70-dim feature redesign and the element KG fusion branch. The 0.005 eV difference in Table 4 (0.177 vs 0.182) is the only support for the KG branch itself, and it is within plausible run-to-run noise given that no error bars or multiple seeds are reported. To support the central claim, the authors should report (i) iComFormer with the 70-dim features but no KG branch, (ii) ESNet with the original 92-dim features plus the KG branch, and (iii) results over multiple seeds with standard deviations.
  3. [§4.1, Table 2] The formation-energy result for ESNet is 20.05 meV/atom, which is worse than both iComFormer (18.26) and PotNet (18.8) despite the abstract and conclusion claiming results 'on a par with existing benchmarks.' This is a misleading characterization of the reported numbers; the narrative should be revised to state plainly that ESNet underperforms the best structural baselines on formation energy while improving band-gap MAE. Additionally, the table labels the proposed method 'SENet' instead of 'ESNet', which is inconsistent with the rest of the manuscript.
  4. [§4.1 and Table 3] The paper reports a single fixed data split (from Matformer and PotNet) and no measure of variance. Since the core claim rests on a small band-gap improvement and the fusion encoder depth differs between tasks (n=8 for band gap and n=4 for formation energy in Table 3), it is unclear whether the observed gain comes from the element KG branch, from the deeper fusion Transformer, or from luck on a single split. Please provide repeated-seed statistics or at least error bars for the main comparisons, and discuss the sensitivity to the fusion encoder depth.
minor comments (5)
  1. [Abstract] The phrase 'achieved results on a par with existing benchmarks' overstates the reported formation-energy numbers relative to iComFormer and PotNet; suggest rewording to match the actual table values.
  2. [§3.3, Figure 1] The figure caption mentions 'The number of modules n in the Fusion Encoder is set according to the specific task,' but the relationship between n and the model's performance is not discussed; please clarify how n was chosen.
  3. [§4.1, text before Table 2] The sentence 'The the best experimental results highlighted in bold' contains a duplicated article; please fix the typo.
  4. [Table 1] The table header 'First ionization energy' is split across two lines in a way that could confuse readers; also, the unit range '3-65' for ionization energy appears to be in eV, but this is not clearly marked; please make the units explicit.
  5. [§3.2, Figure 6] The figure reports validation MAE for the 70- versus 92-dimensional features, but the experimental details (number of epochs, early stopping criteria, and whether the same random seed was used) are omitted; please provide these details for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the element KG is built from independent periodic-table attributes and unsupervised embeddings; the target properties play no role in constructing the input features.

full rationale

The element knowledge graph is constructed from periodic-table attribute triples (e.g., density, electronegativity, atomic radius) and embedded with OWL2Vec/Word2Vec trained on structural, lexical, and combinatorial documents derived from the KG itself. The target properties (band gap, formation energy) never enter KG construction, embedding, or the stoichiometric weighted-sum pooling that produces He. The crystal graph encoder is iComFormer, an external baseline, and the fusion Hf = αHe + βHg is a learned combination rather than a fit to the test labels. The ablation in Table 4 removes the element KG encoder while keeping the redesigned 70-dim atom features; this is a confounded experimental comparison (the 0.005 eV difference may reflect noise or the encoder change), but a confound is not circularity. The paper also relies on external prior work (KCL, KANO, OWL2Vec, iComFormer) rather than a self-citation chain. Acknowledged limitations, such as untested shear/bulk modulus and planned cross-validation, are scope statements rather than circular moves. Therefore, no claimed derivation or prediction reduces by construction to its own inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The pipeline depends on several modeling choices not derived from first principles: the set of element attributes, their discretization, the embedding method, and the fusion weights. These are inputs to the system, not consequences of the physics, and the paper provides no sensitivity analysis.

free parameters (6)
  • fusion weights α and β in Hf = αHe + βHg = not reported
    Control the contribution of element and structure features; values not published, may be learned or tuned per task.
  • fusion encoder depth n = 8 (band gap), 4 (formation energy)
    Table 3, chosen per task.
  • learning rate = 0.0005 (band gap), 0.0007 (formation energy)
    Table 3, hyperparameter.
  • atomic feature discretization for continuous properties = 10 uniform bins per property (e.g., covalent radius 32-232 pm, ionization energy 3-65 eV)
    Section 3.2, Table 1; bin boundaries are modeling choices.
  • element KG attribute discretization intervals = unspecified
    Section 3.1.1 divides continuous attributes into intervals; the chosen boundaries are not reported.
  • selection of six atom properties for the 70-dim vector = 6 properties
    Chosen and validated on the MP validation set (Figure 6), a model selection choice.
assumptions (5)
  • domain assumption MP DFT-computed band gaps and formation energies are reliable ground truth.
    Section 4.1 uses the MP-2018.6.1 dataset labels as targets without uncertainty.
  • domain assumption OWL2Vec/Word2Vec embeddings of an element-attribute KG capture chemically meaningful element similarities.
    Section 3.1.2 assumes semantic embeddings from ontology random walks transfer to property prediction.
  • domain assumption The stoichiometric weighted sum of element embedding vectors represents a crystal's elemental composition adequately.
    Section 3.1.2, Example NaAlSi3O8; this representation discards structure and is assumed sufficient for the feature fusion.
  • domain assumption The 70-dim atomic feature vector preserves all relevant information in the original 92-dim CGCNN descriptors while fixing errors.
    Section 3.2; validated only on the same benchmark's validation set.
  • domain assumption Hyperparameters tuned on the validation split transfer to the test split.
    Section 4.1, hyperparameters in Table 3 chosen per task; standard assumption but not cross-validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Material Property Prediction with Element Attribute Knowledge Graphs and Multimodal Representation Learning." pith.science (2026). https://pith.science/paper/267J7QQX

@misc{pith2026241108414,
  author       = {Pith},
  title        = {Pith review of: Material Property Prediction with Element Attribute Knowledge Graphs and Multimodal Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/267J7QQX}},
  note         = {Machine review of arXiv:2411.08414}
}
read the original abstract

Machine learning has become a crucial tool for predicting the properties of crystalline materials. However, existing methods primarily represent material information by constructing multi-edge graphs of crystal structures, often overlooking the chemical and physical properties of elements (such as atomic radius, electronegativity, melting point, and ionization energy), which have a significant impact on material performance. To address this limitation, we first constructed an element property knowledge graph and utilized an embedding model to encode the element attributes within the knowledge graph. Furthermore, we propose a multimodal fusion framework, ESNet, which integrates element property features with crystal structure features to generate joint multimodal representations. This provides a more comprehensive perspective for predicting the performance of crystalline materials, enabling the model to consider both microstructural composition and chemical characteristics of the materials. We conducted experiments on the Materials Project benchmark dataset, which showed leading performance in the bandgap prediction task and achieved results on a par with existing benchmarks in the formation energy prediction task.

Figures

Figures reproduced from arXiv: 2411.08414 by the authors.

Figure 1
Figure 1. Overview of our adopted methodology ESNet.The complete element embedding is obtained by weighting and summing the individual element vectors output by the Element KG Encoder based on the elemental ratios in the crystal structure.The number of modules n in the Fusion Encoder is set according to the specific task. 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Examples of Fe element related attributes 3.1.1 Element Knowledge Graph Construction In our work, all chemical elements and attributes are obtained from the periodic table (https://ptable.com) , each element contains at least 15 attributes(e.g [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Overview of ElementKG Construction And Embedding. a. Construction of the Elemen￾tal Knowledge Graph. We collect element attribute knowledge from the periodic table and construct element triples. b.Embedding process for elemental knowledge graph. Three docu￾ments (structural, lexical and combinatorial) are derived from the elemental knowledge graph and merged into a single document for training the word embedding mod… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Illustration of the iComFormer pipeline. a. Materialcrystal structure. b. Crystal multimodal graph. c. iComFormer architecture, where elements of the same color belong to the same module. Specifically, we redesigned the atomic encoding in the crystal graph, reducing th…
Figure 5
Figure 5. Figure 5: "H" atoms vector in CGCNN. The blue dashed boxes in the figure are used to distinguish other elements from the lanthanide and actinide elements. There are the following issues with the encoding of the hydrogen atom "H": a. The electronegativity is 2.2, so the value in …
Figure 6
Figure 6. Figure 6: The results on the Materials Project validation set, where both the 70-dimensional and 92- dimensional atomic feature vectors were trained using the same parameters. a. The MAE results for the band gap on the validation set: using L1 loss with Adam optimizer and a star…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 23 canonical work pages

  1. [1]

    Sankaranarayanan

    Suvo Banik, Karthik Balasubramanian, Sukriti Manna, Sybil Derrible, and Subramanian K.R. Sankaranarayanan. Evaluating generalized feature importance via performance assessment of machine learning models for predicting elastic properties of materials. Computational Materials Science, 220: 0 110687, 2024. doi:https://doi.org/10.1016/j.commatsci.2023.110687

    work page arXiv 2024

  2. [2]

    Graph networks as a universal machine learning framework for molecules and crystals

    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 0 (9): 0 3564--3572, 2019

    work page 2019

  3. [3]

    Owl2vec*: Embedding of owl ontologies

    Jiaoyan Chen, Pan Hu, Ernesto Jimenez-Ruiz, Ole Magnus Holter, Denvar Antonyrajah, and Ian Horrocks. Owl2vec*: Embedding of owl ontologies. Machine Learning, 110 0 (7): 0 1813--1845, 2021

    work page 2021

  4. [4]

    Atomistic line graph neural network for improved materials property predictions

    Kamal Choudhary and Brian DeCost. Atomistic line graph neural network for improved materials property predictions. npj Computational Materials, 7 0 (1): 0 185, 2021 a . doi:10.1038/s41524-021-00650-1

    doi:10.1038/s41524-021-00650-1 2021

  5. [5]

    Atomistic line graph neural network for improved materials property predictions

    Kamal Choudhary and Brian DeCost. Atomistic line graph neural network for improved materials property predictions. npj Computational Materials, 7 0 (1): 0 185, 2021 b

    work page 2021

  6. [6]

    Atomistic line graph neural network for improved materials property predictions

    DeCost B Choudhary K. Atomistic line graph neural network for improved materials property predictions. npj Computational Materials, 7 0 (1): 0 1--8, 2021

    work page 2021

  7. [7]

    Crysmmnet: multimodal representation for crystal property prediction

    Kishalay Das, Pawan Goyal, Seung-Cheol Lee, Satadeep Bhattacharjee, and Niloy Ganguly. Crysmmnet: multimodal representation for crystal property prediction. In Uncertainty in Artificial Intelligence, pages 507--517. PMLR, 2023

    work page 2023

  8. [8]

    Molecular contrastive learning with chemical element knowledge graph

    Yin Fang, Qiang Zhang, Haihong Yang, Xiang Zhuang, Shumin Deng, Wen Zhang, Ming Qin, Zhuo Chen, Xiaohui Fan, and Huajun Chen. Molecular contrastive learning with chemical element knowledge graph. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 3968--3976, 2022

    work page 2022

Show all 31 references
  1. [9]

    Knowledge graph-enhanced molecular contrastive learning with functional prompt

    Yin Fang, Qiang Zhang, Ningyu Zhang, Zhuo Chen, Xiang Zhuang, Xin Shao, Xiaohui Fan, and Huajun Chen. Knowledge graph-enhanced molecular contrastive learning with functional prompt. Nature Machine Intelligence, 5 0 (5): 0 542--553, 2023

    2023

  2. [10]

    Robocrystallographer: automated crystal structure text descriptions and analysis

    Alex M Ganose and Anubhav Jain. Robocrystallographer: automated crystal structure text descriptions and analysis. MRS Communications, 9 0 (3): 0 874--881, 2019

    2019

  3. [11]

    word2vec, node2vec, graph2vec, x2vec: Towards a theory of vector embeddings of structured data

    Martin Grohe. word2vec, node2vec, graph2vec, x2vec: Towards a theory of vector embeddings of structured data. In Proceedings of the 39th ACM SIGMOD-SIGACT-SIGAI Symposium on Principles of Database Systems, pages 1--16, 2020

    2020

  4. [12]

    Machine learning based feature engineering for thermoelectric materials by design

    William L Hamilton, Zhitao Ying, and Jure Leskovec. Machine learning based feature engineering for thermoelectric materials by design. Digital Discovery, 3: 0 210--220, 2024. doi:10.1039/D3DD00131H

    2024 doi

  5. [13]

    Hwang, J

    S. Hwang, J. Jung, and C. Hong. Stability and equilibrium structures of unknown ternary metal oxides explored by machine-learned potentials. J. Am. Chem. Soc., 2023

    2023

  6. [14]

    Universal fragment descriptors for predicting properties of inorganic crystals

    Olexandr Isayev, Carlos Oses, Celine Toher, Elizabeth Gossett, and Stefano Curtarolo. Universal fragment descriptors for predicting properties of inorganic crystals. Nature Communications, 8: 0 15679, 2017. doi:10.1038/ncomms15679. URL https://www.nature.com/articles/ncomms15679

    2017 doi

  7. [15]

    Materials representation and transfer learning for multi-property prediction

    Shufeng Kong, Dan Guevarra, Carla P Gomes, and John M Gregoire. Materials representation and transfer learning for multi-property prediction. Applied Physics Reviews, 8 0 (2), 2021

    2021

  8. [16]

    Efficient approximations of complete interatomic potentials for crystal property prediction

    Yuchao Lin, Keqiang Yan, Youzhi Luo, Yi Liu, Xiaoning Qian, and Shuiwang Ji. Efficient approximations of complete interatomic potentials for crystal property prediction. In International Conference on Machine Learning, pages 21260--21287. PMLR, 2023

    2023

  9. [17]

    Graph convolutional neural networks with global attention for improved materials property prediction

    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 0 (32): 0 18141--18148, 2020

    2020

  10. [18]

    Nanomine: A knowledge graph for nanocomposite materials science

    Jamie P McCusker, Neha Keshan, Sabbir Rashid, Michael Deagen, Cate Brinson, and Deborah L McGuinness. Nanomine: A knowledge graph for nanocomposite materials science. In International Semantic Web Conference, pages 144--159. Springer, 2020

    2020

  11. [19]

    Machine learning in materials science: Recent progress and emerging applications

    Tim Mueller, Aaron Gilad Kusne, and Rampi Ramprasad. Machine learning in materials science: Recent progress and emerging applications. Reviews in computational chemistry, 29: 0 186--273, 2016

    2016

  12. [20]

    Anubhav Jain; Shyue Ping Ong; Geoffroy Hautier; Wei Chen; William Davidson Richards; Stephen Dacek; Shreyas Cholia; Dan Gunter; David Skinner; Gerbrand Ceder; Kristin A. Persson. Commentary: The materials project: A materials genome approach to accelerating materials innovatio...

    2013

  13. [21]

    u tt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert M \

    Kristof Sch \"u tt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert M \"u ller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. Advances in neural information processing system...

    2017

  14. [22]

    Matkg: The largest knowledge graph in materials science--entities, relations, and link prediction through graph representation learning

    Vineeth Venugopal, Sumit Pai, and Elsa Olivetti. Matkg: The largest knowledge graph in materials science--entities, relations, and link prediction through graph representation learning. arXiv preprint arXiv:2210.17340, 2022

    2022 arXiv

  15. [23]

    Grossman

    Tian Xie and Jeffrey C. Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Physical Review Letters, 120 0 (14): 0 145301, 2018 a . doi:10.1103/PhysRevLett.120.145301

    2018 doi

  16. [24]

    Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties

    Tian Xie and Jeffrey C Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Physical review letters, 120 0 (14): 0 145301, 2018 b

    2018

  17. [25]

    Periodic graph transformers for crystal material property prediction

    Keqiang Yan, Yi Liu, Yuchao Lin, and Shuiwang Ji. Periodic graph transformers for crystal material property prediction. Advances in Neural Information Processing Systems, 35: 0 15066--15080, 2022 a

    2022

  18. [26]

    Periodic graph transformers for crystal material property prediction

    Keqiang Yan, Yi Liu, Yuchao Lin, and Shuiwang Ji. Periodic graph transformers for crystal material property prediction. Advances in Neural Information Processing Systems, 35: 0 15066--15080, 2022 b

    2022

  19. [28]

    Complete and efficient graph transformers for crystal material property prediction

    Keqiang Yan, Cong Fu, Xiaofeng Qian, Xiaoning Qian, and Shuiwang Ji. Complete and efficient graph transformers for crystal material property prediction. arXiv preprint arXiv:2403.11857, 2024 b

    2024 arXiv

  20. [29]

    Artificial intelligence for science in quantum, atomistic, and continuum systems

    Xuan Zhang, Limei Wang, Jacob Helwig, Youzhi Luo, Cong Fu, Yaochen Xie, Meng Liu, Yuchao Lin, Zhao Xu, Keqiang Yan, et al. Artificial intelligence for science in quantum, atomistic, and continuum systems. arXiv preprint arXiv:2307.08423, 2023

    2023 arXiv

  21. [30]

    Machine-learning-assisted multiscale modeling strategy for predicting mechanical properties of carbon fiber reinforced polymers

    Guomei Zhao, Tianhao Xu, Xuemeng Fu, Wenlin Zhao, Liquan Wang, Jiaping Lin, Yaxi Hu, and Lei Du. Machine-learning-assisted multiscale modeling strategy for predicting mechanical properties of carbon fiber reinforced polymers. Composites Science and Technology, 226: 0 108849, 2024

    2024

  22. [31]

    High-throughput computing assisted by knowledge graph to study the correlation between microstructure and mechanical properties of 6xxx aluminum alloy

    Xiaoyu Zheng, Yi Kong, Tingting Chang, Xin Liao, Yiwu Ma, and Yong Du. High-throughput computing assisted by knowledge graph to study the correlation between microstructure and mechanical properties of 6xxx aluminum alloy. Materials, 15 0 (15): 0 5296, 2022

    2022

  23. [32]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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