pith. sign in

arxiv: 2512.06697 · v2 · submitted 2025-12-07 · ❄️ cond-mat.mtrl-sci · physics.app-ph

Learning Thermoelectric Transport from Crystal Structures via Multiscale Graph Neural Network

Yuxuan Zeng , Wei Cao , Yijing Zuo , Fang Lyu , Wenhao Xie , Tan Peng , Yue Hou , Ling Miao
show 2 more authors
Ziyu Wang Jing Shi
This is my paper

Pith reviewed 2026-05-17 01:20 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci physics.app-ph
keywords thermoelectric transportgraph neural networkcrystal structureselectronic transport coefficientsmultiscale encodingmaterials discoveryinterpretability analysis
0
0 comments X

The pith

A multiscale graph neural network estimates electronic transport coefficients in thermoelectric crystals directly from their structures.

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

The paper develops a graph neural network that encodes crystal structures and related properties at global, atomic, bond, and angular scales to predict electronic transport behavior in inorganic thermoelectric materials. This matters because finding materials that efficiently convert heat to electricity has relied on slow, expensive calculations for each candidate. If the approach holds, it can screen far more structures quickly, flag promising compounds for deeper study, and surface physical patterns that guide future design. The model is tested for accuracy on known benchmarks and for its ability to generalize to unseen crystals. When the predictions are paired with first-principles calculations, the work identifies real compounds that show strong transport performance and uses interpretability tools to link specific structural features to the observed behavior.

Core claim

Encoding crystal structures and physicochemical properties in a multiscale manner, encompassing global, atomic, bond, and angular levels, allows a graph neural network to estimate electronic transport coefficients in inorganic thermoelectric crystals, achieving state-of-the-art performance on benchmark datasets with strong extrapolative capability; when combined with ab initio calculations, the model identifies compounds with outstanding electronic transport properties and interpretability analyses from global and atomic perspectives trace the origins of their distinct transport behaviors, with the model's decision process revealing underlying physical patterns.

What carries the argument

Multiscale graph neural network encoding that integrates global crystal features with atomic, bond, and angular information to represent structural influences on transport.

If this is right

  • Compounds with outstanding electronic transport properties can be identified by pairing the model with ab initio calculations.
  • Interpretability analyses from global and atomic perspectives can trace the structural origins of distinct transport behaviors.
  • The model's decision process can reveal underlying physical patterns that inform materials design.
  • State-of-the-art accuracy and extrapolative capability on benchmark datasets follow from the multiscale encoding.

Where Pith is reading between the lines

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

  • The same multiscale encoding strategy could be tested on predictions of thermal conductivity or other transport-related properties.
  • Larger training sets spanning more crystal families might further strengthen extrapolation to entirely novel structures.
  • The revealed physical patterns could be checked against independent experimental measurements on the highlighted compounds.

Load-bearing premise

That encoding crystal structures and physicochemical properties at global, atomic, bond, and angular levels captures the underlying physics of electronic transport coefficients.

What would settle it

Ab initio calculations or measurements on the newly identified high-transport compounds that yield electronic transport values substantially different from the model's predictions would falsify the performance and discovery claims.

Figures

Figures reproduced from arXiv: 2512.06697 by Fang Lyu, Jing Shi, Ling Miao, Tan Peng, Wei Cao, Wenhao Xie, Yijing Zuo, Yue Hou, Yuxuan Zeng, Ziyu Wang.

Figure 1
Figure 1. Figure 1: FIG. 1. Graph representation of crystal structures. Tak [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. The proposed GNN framework for crystal struc [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Error evaluation for the TECSA-GNN model. (a)-(c) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. 10-fold cross-validation performance of the TECSA [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Parity plots of TECSA-GNN after fine-tuning on special structures. (a) [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Comparison between TE transport properties pre [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Projected band structure and PDOS. (a) Te [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Atomic importance and ELF spatial visualization. In the unit cells of (a) Te [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Partial dependence plots of the key global features [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
read the original abstract

Graph neural networks (GNNs) are designed to extract latent patterns from graph-structured data, making them particularly well suited for crystal representation learning. Here, we propose a GNN model tailored for estimating electronic transport coefficients in inorganic thermoelectric crystals. The model encodes crystal structures and physicochemical properties in a multiscale manner, encompassing global, atomic, bond, and angular levels. It achieves state-of-the-art performance on benchmark datasets with remarkable extrapolative capability. By combining the proposed GNN with \textit{ab initio} calculations, we successfully identify compounds exhibiting outstanding electronic transport properties and further perform interpretability analyses from both global and atomic perspectives, tracing the origins of their distinct transport behaviors. Interestingly, the decision process of the model naturally reveals underlying physical patterns, offering new insights into computer-assisted materials design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The manuscript proposes a multiscale graph neural network (GNN) for predicting electronic transport coefficients in inorganic thermoelectric crystals. The model encodes structures and properties at global, atomic, bond, and angular levels, claiming state-of-the-art performance on benchmark datasets together with remarkable extrapolative capability. The GNN is combined with ab initio calculations to identify new compounds with outstanding transport properties, and interpretability analyses are presented to trace physical origins of the predictions.

Significance. If the extrapolative claims hold under proper out-of-distribution testing, the work could accelerate screening for high-performance thermoelectrics and supply interpretable physical insights that complement traditional band-structure calculations.

major comments (3)
  1. [§4 (Benchmark Results)] §4 (Benchmark Results): The SOTA performance and extrapolative capability are asserted without reporting standard deviations across multiple runs, statistical significance tests versus baselines, or explicit description of the train/validation/test splits (random, scaffold, or elemental OOD). Thermoelectric coefficients depend sensitively on chemistry-specific band details, so the absence of these details leaves the central generalization claim unsupported.
  2. [§5 (Compound Discovery)] §5 (Compound Discovery): The pipeline that uses GNN predictions to select candidates for ab initio validation is load-bearing for the discovery claim, yet the selection criteria, number of screened versus validated compounds, and any false-positive rate are not quantified. This makes it impossible to judge whether the extrapolative step actually yields reliable new materials.
  3. [Model Architecture (Methods)] Model Architecture (Methods): The fusion of global, atomic, bond, and angular encodings is described at a high level but lacks an explicit equation or pseudocode showing how angular features are computed and combined with other scales. Without this, it is difficult to assess whether the multiscale design genuinely captures the physics needed for transport coefficients.
minor comments (3)
  1. [Abstract] Abstract: The phrase 'remarkable extrapolative capability' should be accompanied by a concrete metric or example drawn from the results section.
  2. [Figures] Figures: Performance plots and interpretability visualizations would benefit from consistent axis labels, legends, and inclusion of error bars where applicable.
  3. [References] References: Ensure all benchmark datasets and prior GNN works for materials are cited with their original sources.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important aspects of statistical rigor, pipeline transparency, and architectural clarity that we address point by point below. We have revised the manuscript to incorporate additional details and clarifications where they strengthen the presentation without altering the core claims or results.

read point-by-point responses

  1. Referee: [§4 (Benchmark Results)] §4 (Benchmark Results): The SOTA performance and extrapolative capability are asserted without reporting standard deviations across multiple runs, statistical significance tests versus baselines, or explicit description of the train/validation/test splits (random, scaffold, or elemental OOD). Thermoelectric coefficients depend sensitively on chemistry-specific band details, so the absence of these details leaves the central generalization claim unsupported.

    Authors: We agree that explicit reporting of variability and split details is necessary to substantiate the generalization claims. Although experiments were run with multiple random seeds during development, the standard deviations and formal significance tests were not included in the original main text. In the revised version we have added these to Table 2 (mean and standard deviation over five independent runs) together with paired t-test p-values against each baseline. We have also expanded Section 3.2 to describe the data partitioning explicitly: primary results use random splits stratified by composition, while supplementary results include elemental out-of-distribution splits that withhold entire chemical families. These additions directly address the concern regarding chemistry-specific band details and the robustness of the extrapolative claims. revision: yes

  2. Referee: [§5 (Compound Discovery)] §5 (Compound Discovery): The pipeline that uses GNN predictions to select candidates for ab initio validation is load-bearing for the discovery claim, yet the selection criteria, number of screened versus validated compounds, and any false-positive rate are not quantified. This makes it impossible to judge whether the extrapolative step actually yields reliable new materials.

    Authors: We acknowledge that a quantitative description of the screening pipeline is essential for evaluating the reliability of the discovered compounds. The original manuscript summarized the outcome but did not tabulate the full workflow. In the revised Section 5 we now specify the selection criteria (GNN-predicted electronic figure of merit above a threshold, combined with stability filters), report the total number of structures screened from the candidate pool, the subset advanced to DFT validation, and the agreement rate between GNN predictions and ab initio results for the validated set. This provides a transparent basis for assessing the false-positive behavior of the extrapolative step. revision: yes

  3. Referee: [Model Architecture (Methods)] Model Architecture (Methods): The fusion of global, atomic, bond, and angular encodings is described at a high level but lacks an explicit equation or pseudocode showing how angular features are computed and combined with other scales. Without this, it is difficult to assess whether the multiscale design genuinely captures the physics needed for transport coefficients.

    Authors: We concur that an explicit formulation improves both reproducibility and physical interpretability. The revised Methods section now includes Equation (3), which defines the angular feature vector via a learned projection of bond-angle cosine terms and spherical-harmonic expansions up to order l=2, and Algorithm 1, which shows the concatenation and gated fusion of the four scale-specific embeddings before the final readout. These additions make clear how angular information is integrated with global, atomic, and bond representations to capture the directional dependencies relevant to electronic transport. revision: yes

Circularity Check

0 steps flagged

No significant circularity; model training and external validation are independent

full rationale

The paper trains a multiscale GNN on external benchmark datasets for thermoelectric transport coefficients and evaluates performance on held-out test portions of those benchmarks. New candidate compounds are then proposed via model inference and independently verified or refined with ab initio calculations. No equation or claim reduces by construction to the inputs (e.g., no fitted parameter is relabeled as a prediction, no self-citation chain substitutes for a derivation, and no ansatz is smuggled via prior work). The pipeline remains self-contained against external data sources and first-principles follow-up, satisfying the default expectation of non-circularity for empirical ML papers.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard graph representations of crystals and numerous fitted neural-network parameters; no new physical entities are postulated.

free parameters (1)
  • GNN hyperparameters and training parameters
    Model architecture choices and optimization parameters are tuned on benchmark data.
axioms (1)
  • domain assumption Crystal structures can be faithfully represented as multiscale graphs with nodes, edges, and angular features.
    Invoked in the model design section implied by the abstract.

pith-pipeline@v0.9.0 · 6690 in / 1150 out tokens · 82104 ms · 2026-05-17T01:20:53.997822+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

104 extracted references · 104 canonical work pages

  1. [1]

    He and T

    J. He and T. M. Tritt, Science357, eaak9997 (2017)

    work page 2017

  2. [2]

    X. Sun, Y. Hou, Z. Zhu, B. Zhu, Q. Ding, W. Zhou, S. Yan, Z. Xia, Y. Liu, Y. Hou,et al., Nat. Commun. 16, 1 (2025)

    work page 2025

  3. [3]

    W. Yang, Y. Zhong, D. Ao, D. Yang, M. Wei, F. Li, Y. Chen, G. Liang, J. Luo, and Z. Zheng, Appl. Phys. Lett.126, 123902 (2025)

    work page 2025

  4. [4]

    Kraemer, Q

    D. Kraemer, Q. Jie, K. McEnaney, F. Cao, W. Liu, L. A. Weinstein, J. Loomis, Z. Ren, and G. Chen, Nat. Energy 1, 1 (2016)

    work page 2016

  5. [5]

    T. Y. Kim, C.-H. Park, and N. Marzari, Nano Lett.16, 2439 (2016)

    work page 2016

  6. [6]

    J.Moi, D.Spallarossa, S.Bonetti, R.Burioni,andG.Cal- darelli, Phys. Rev. E112, 011001 (2025)

    work page 2025

  7. [7]

    Leeman, Y

    J. Leeman, Y. Liu, J. Stiles, S. B. Lee, P. Bhatt, L. M. Schoop, and R. G. Palgrave, PRX Energy3, 011002 (2024)

    work page 2024

  8. [8]

    L. M. Antunes, K. T. Butler, and R. Grau-Crespo, Mach. Learn.: Sci. Technol.4, 015037 (2023)

    work page 2023

  9. [9]

    Y. Xu, L. Jiang, and X. Qi, Comput. Mater. Sci.197, 110625 (2021)

    work page 2021

  10. [10]

    Al-Fahdi, K

    M. Al-Fahdi, K. Yuan, Y. Yao, R. Rurali, and M. Hu, Appl. Phys. Rev.11, 021402 (2024)

    work page 2024

  11. [11]

    J. Yan, P. Gorai, B. Ortiz, S. Miller, S. A. Barnett, T. Mason, V. Stevanović, and E. S. Toberer, Energy Environ. Sci.8, 983 (2015)

    work page 2015

  12. [12]

    M. Y. Toriyama, A. N. Carranco, G. J. Snyder, and P. Gorai, Mater. Horiz.10, 4256 (2023)

    work page 2023

  13. [13]

    Huang, S

    X. Huang, S. Ma, C. Zhao, H. Wang, and S. Ju, npj Comput. Mater.9, 191 (2023)

    work page 2023

  14. [14]

    W. Cao, J. Shi, R. Xiong, L. Miao, Z. Wang, and Z. Liu, Phys. Rev. B107, 235201 (2023)

    work page 2023

  15. [15]

    F. Lyu, W. Cao, H.-P. Liang, T. Peng, Y. Hou, X. Zhu, L. Miao, Z. Wang, R. Xiong, and J. Shi, J. Mater. Chem. A12, 18452 (2024)

    work page 2024

  16. [16]

    Scarselli, M

    F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini, IEEE Trans. Neural Netw.20, 61 (2008)

    work page 2008

  17. [17]

    Y. Wang, J. Wang, Z. Cao, and A. Barati Farimani, Nat. Mach. Intell.4, 279 (2022)

    work page 2022

  18. [18]

    C. W. Coley, R. Barzilay, W. H. Green, T. S. Jaakkola, and K. F. Jensen, J. Chem. Inf. Model.57, 1757 (2017)

    work page 2017

  19. [19]

    M Veselinovic, J

    A. M Veselinovic, J. B Veselinovic, J. V Zivkovic, and G. M Nikolic, Curr. Top. Med. Chem.15, 1768 (2015)

    work page 2015

  20. [20]

    Xie and J

    T. Xie and J. C. Grossman, Phys. Rev. Lett.120, 145301 (2018)

    work page 2018

  21. [21]

    Choubisa, P

    H. Choubisa, P. Todorović, J. M. Pina, D. H. Parmar, Z. Li, O. Voznyy, I. Tamblyn, and E. H. Sargent, npj Comput. Mater.9, 117 (2023)

    work page 2023

  22. [22]

    Choudhary and B

    K. Choudhary and B. DeCost, npj Comput. Mater.7, 185 (2021)

    work page 2021

  23. [23]

    Karamad, R

    M. Karamad, R. Magar, Y. Shi, S. Siahrostami, I. D. Gates, and A. Barati Farimani, Phys. Rev. Mater.4, 093801 (2020)

    work page 2020

  24. [24]

    C. W. Park and C. Wolverton, Phys. Rev. Mater.4, 063801 (2020)

    work page 2020

  25. [25]

    Z. Wang, R. Hu, X. Luo, and J. Ma, J. Appl. Phys.136, 155103 (2024)

    work page 2024

  26. [26]

    Gómez and A

    D. Gómez and A. Rojas, Neural Comput.28, 216 (2016)

    work page 2016

  27. [27]

    E. Lin, Y. Zhong, G. Chen, and S. Deng, npj Comput. Mater.11, 225 (2025)

    work page 2025

  28. [28]

    L. Dong, X. Zhang, Z. Yang, L. Shen, and Y. Lu, npj Comput. Mater.11, 63 (2025)

    work page 2025

  29. [29]

    Ricci, W

    F. Ricci, W. Chen, U. Aydemir, G. J. Snyder, G.-M. Rignanese, A. Jain, and G. Hautier, Sci. Data4, 1 (2017)

    work page 2017

  30. [30]

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, IEEE Trans. Neural Netw. Learn. Syst.32, 4 (2020)

    work page 2020

  31. [31]

    Vipin and P

    K. Vipin and P. Padhan, J. Mater. Chem. C12, 7415 (2024)

    work page 2024

  32. [32]

    Choubisa, M

    H. Choubisa, M. A. Haque, T. Zhu, L. Zeng, M. Vafaie, D. Baran, and E. H. Sargent, Adv. Mater.35, 2302575 (2023)

    work page 2023

  33. [33]

    Niepert, M

    M. Niepert, M. Ahmed, and K. Kutzkov, inInt. Conf. Mach. Learn., Proc. Mach. Learn. Res., Vol. 48 (PMLR, New York, New York, USA, 2016) pp. 2014–2023

    work page 2016

  34. [34]

    W. G. Zeier, A. Zevalkink, Z. M. Gibbs, G. Hautier, M. G. Kanatzidis, and G. J. Snyder, Angew. Chem. Int. Ed.55, 6826 (2016)

    work page 2016

  35. [35]

    F. M. A. Acosta, Signal Proc.45, 37 (1995)

    work page 1995

  36. [36]

    Veličković, G

    P. Veličković, G. Cucurull, A. Casanova, A. Romero, P. Liò, and Y. Bengio, inInt. Conf. Learn. Represent. (2018)

    work page 2018

  37. [37]

    S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder, Comput. Mater. Sci.68, 314 (2013). 15

    work page 2013

  38. [38]

    J. P. Higgins, I. R. White, and J. Anzures-Cabrera, Stat. Med.27, 6072 (2008)

    work page 2008

  39. [39]

    A. Y.-T. Wang, S. K. Kauwe, R. J. Murdock, and T. D. Sparks, npj Comput. Mater.7, 77 (2021)

    work page 2021

  40. [40]

    S. J. Pan and Q. Yang, IEEE Trans. Knowl. Data Eng. 22, 1345 (2010)

    work page 2010

  41. [41]

    Mingard, H

    C. Mingard, H. Rees, G. Valle-Pérez, and A. A. Louis, Nat. Commun.16, 220 (2025)

    work page 2025

  42. [42]

    J. Li, Z. Chen, J. Wu, J. Lin, P. He, R. Zhu, C. Peng, H. Zhang, W. Li, X. Fang, and H. Shen, Mater. Today Commun.35, 106299 (2023)

    work page 2023

  43. [43]

    Dhahri, A

    R. Dhahri, A. Immer, B. Charpentier, S. Günnemann, and V. Fortuin, inAdv. Neural Inform. Process. Syst., Vol. 37 (Curran Associates, Inc., 2024) pp. 24959–24989

    work page 2024

  44. [44]

    Cheng, C

    J. Cheng, C. Zhang, and L. Dong, Commun. Mater.2, 92 (2021)

    work page 2021

  45. [45]

    L. v. d. Maaten and G. Hinton, J. Mach. Learn. Res.9, 2579 (2008)

    work page 2008

  46. [46]

    Zhang, P

    H. Zhang, P. Wang, X. Gao, Y. Qi, and H. Gao, Inf. Visualization20, 20 (2021)

    work page 2021

  47. [47]

    Foster and N

    S. Foster and N. Neophytou, Comput. Mater. Sci.164, 91 (2019)

    work page 2019

  48. [48]

    Graziosi, Z

    P. Graziosi, Z. Li, and N. Neophytou, Appl. Phys. Lett. 120, 072102 (2022)

    work page 2022

  49. [49]

    M. D. McKay, M. A. Fitzgerald, and R. J. Beckman (Los Alamos National Lab., NM (US), 1999)

    work page 1999

  50. [50]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, et al., APL Mater.1, 10.1063/1.4812323 (2013)

    work page doi:10.1063/1.4812323 2013

  51. [51]

    B. Su, Z. Han, Y. Jiang, H.-L. Zhuang, J. Yu, J. Pei, H. Hu, J.-W. Li, Y.-X. He, B.-P. Zhang,et al., Adv. Funct. Mater.33, 2301971 (2023)

    work page 2023

  52. [52]

    Yu, X.-L

    L. Yu, X.-L. Shi, Y. Mao, M. Li, W.-D. Liu, Z. Ji, S. Wei, Z. Zhang, W. Song, S. Zheng, and Z.-G. Chen, Chem. Eng. J.482, 149051 (2024)

    work page 2024

  53. [53]

    Q. Xu, J. Zhou, T.-H. Liu, and G. Chen, Appl. Phys. Lett.115, 023903 (2019)

    work page 2019

  54. [54]

    C. Wan, Y. Wang, N. Wang, W. Norimatsu, M. Kusunoki, and K. Koumoto, Sci. Technol. Adv. Mater. 11, 044306 (2010)

    work page 2010

  55. [55]

    J. M. Buhmann and M. Sigrist, Phys. Rev. B88, 115128 (2013)

    work page 2013

  56. [56]

    J. Park, Y. Xia, V. Ozolin,š, and A. Jain, npj Comput. Mater.7, 43 (2021)

    work page 2021

  57. [57]

    J. Yang, G. Zhang, G. Yang, C. Wang, and Y. X. Wang, J. Alloys Compd.644, 615 (2015)

    work page 2015

  58. [58]

    Z.-H. Liu, P. Richard, N. Xu, G. Xu, Y. Li, X.-C. Fang, L.-L. Jia, G.-F. Chen, D.-M. Wang, J.-B. He, T. Qian, J.-P. Hu, H. Ding, and S.-C. Wang, Phys. Rev. Lett. 109, 037003 (2012)

    work page 2012

  59. [59]

    D. Wang, Y. Chao, K. Guo, Z. Wang, M. Yang, J. Zhu, X. Cui, and Q. Xu, Adv. Funct. Mater.34, 2405642 (2024)

    work page 2024

  60. [60]

    Gnanapoongothai, R

    T. Gnanapoongothai, R. Murugan, and B. Palanivel, Ionics21, 1351 (2015)

    work page 2015

  61. [61]

    Savin, R

    A. Savin, R. Nesper, S. Wengert, and T. F. Fässler, Angew. Chem. Int. Ed.36, 1808 (1997)

    work page 1997

  62. [62]

    Siddique, G

    S. Siddique, G. Abbas, M. M. Yaqoob, J. Zhao, R. Chen, J. A. Larsson, Y. Cao, Y. Chen, Z. Zheng, D. Zhang, and F. Li, Adv. Sci.12, 2411594 (2025)

    work page 2025

  63. [63]

    B. A. Williamson, J. Buckeridge, J. Brown, S. Ansbro, R. G. Palgrave, and D. O. Scanlon, Chem. Mater.29, 2402 (2017)

    work page 2017

  64. [64]

    W. Zhao, P. Wei, Q. Zhang, H. Peng, W. Zhu, D. Tang, J. Yu, H. Zhou, Z. Liu, and X. Mu, Nat. Commun.6, 6197 (2015)

    work page 2015

  65. [65]

    Jiang, H

    X. Jiang, H. Fu, Y. Bai, L. Jiang, H. Zhang, W. Wang, P. Yun, J. He, D. Xue, T. Lookman,et al., Adv. Funct. Mater. , 2507734 (2025)

    work page 2025

  66. [66]

    Miao and F.-Y

    Q. Miao and F.-Y. Wang, inArtificial Intelligence for Science (AI4S) Frontiers and Perspectives Based on Par- allel Intelligence(Springer, 2024) pp. 105–113

    work page 2024

  67. [67]

    H. Yuan, H. Yu, S. Gui, and S. Ji, IEEE Trans. Pattern Anal. Mach. Intell.45, 5782 (2022)

    work page 2022

  68. [68]

    P. E. Pope, S. Kolouri, M. Rostami, C. E. Martin, and H.Hoffmann,inIEEE Conf. Comput. Vis. Pattern Recog. (2019) pp. 10772–10781

    work page 2019

  69. [69]

    D. Luo, W. Cheng, D. Xu, W. Yu, B. Zong, H. Chen, and X. Zhang, inAdv. Neural Inform. Process. Syst., Vol. 33 (Curran Associates, Inc., 2020) pp. 19620–19631

    work page 2020

  70. [70]

    Schnake, O

    T. Schnake, O. Eberle, J. Lederer, S. Nakajima, K. T. Schütt, K.-R. Müller, and G. Montavon, IEEE Trans. Pattern Anal. Mach. Intell.44, 7581 (2022)

    work page 2022

  71. [71]

    Vu and M

    M. Vu and M. T. Thai, inAdv. Neural Inform. Process. Syst., Vol. 33 (Curran Associates, Inc., 2020) pp. 12225– 12235

    work page 2020

  72. [72]

    Neural Inform

    R.Ying, D.Bourgeois, J.You, M.Zitnik,andJ.Leskovec, GNNExplainer: generating explanations for graph neural networks, inAdv. Neural Inform. Process. Syst.(Curran Associates Inc., Red Hook, NY, USA, 2019)

    work page 2019

  73. [73]

    S. M. Lundberg and S.-I. Lee, Adv. Neural Inform. Pro- cess. Syst.30, 4768–4777 (2017)

    work page 2017

  74. [74]

    J. H. Friedman, Ann. Math. Stat. , 1189 (2001)

    work page 2001

  75. [75]

    B. Russ, A. Glaudell, J. J. Urban, M. L. Chabinyc, and R. A. Segalman, Nat. Rev. Mater.1, 1 (2016)

    work page 2016

  76. [76]

    Fritzsche, Solid State Commun.9, 1813 (1971)

    H. Fritzsche, Solid State Commun.9, 1813 (1971)

    work page 1971

  77. [77]

    J. Gong, A. Hong, J. Shuai, L. Li, Z. Yan, Z. Ren, and J.-M. Liu, Phys. Chem. Chem. Phys.18, 16566 (2016)

    work page 2016

  78. [78]

    J. Qiu, S. Zhi, P. Zhao, J. Wang, X. Ma, S. Ye, C. Lin, X. Zhang, Z. Wu, S. Duan, H. Yao, F. Cao, Q. Zhang, and J. Mao, Phys. Rev. B111, 045203 (2025)

    work page 2025

  79. [79]

    Molnar,Interpretable Machine Learning, 3rd ed

    C. Molnar,Interpretable Machine Learning, 3rd ed. (2025)

    work page 2025

  80. [80]

    T. L. Meek and L. D. Garner, J. Chem. Educ.82, 325 (2005)

    work page 2005

Showing first 80 references.