Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Accelerated Discovery of Vanadium Oxide Compositions: A WGAN-VAE Framework for Materials Design

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A generative WGAN-VAE framework, constrained by formation-energy predictions and validated by DFT plus phonon calculations, discovers new stable vanadium oxide compositions, including two V2O3 phases below the Materials Project convex hull.

desk verdict A competent generative-materials pipeline whose headline below-hull V2O3 claim rests on an unverified reference-frame conversion and DFT tolerances too loose to support it. read the letter →

arxiv 2501.04604 v2 pith:R4YL5H7G submitted 2025-01-08 cond-mat.mtrl-sci cond-mat.dis-nn

Danial Ebrahimzadeh , Sarah S. Sharif , Yaser M. Banad This is my paper · ORCID
classification cond-mat.mtrl-scicond-mat.dis-nn
keywords inversedesignmachinelearningvanadiumoxidesWGANmaterialsdiscoveryelectronicpropertiesDFT+Ucalculationsvariationalautoencoder
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that an inverse-design pipeline can invent new stable vanadium oxide crystals without brute-force screening. The pipeline couples a Wasserstein generative adversarial network (WGAN) that is constrained by formation-energy predictions with a voxel-based variational autoencoder (VAE) that encodes atomic positions and lattice parameters; generated candidates are then relaxed and evaluated with density functional theory. The paper reports 451 unique V–O compositions, classifying 91 as stable and 44 as metastable under criteria stricter than an earlier benchmark, and claims a stability rate of about 20 percent. Its headline discovery is two V2O3 configurations with formation energies below the Materials Project convex hull, which would be previously undocumented thermodynamically favorable phases. Spin-polarized DFT+U calculations on the lowest-energy structures indicate half-metallic behavior in VO2 and V2O3, which would make them relevant for spin-filtered electronics.

What carries the argument

The machinery is a two-stage generative model. A beta-VAE with three 3D voxel grids—two for vanadium and oxygen atomic positions plus one for lattice parameters—compresses crystals into a continuous latent space (a 200-dimensional site vector and a 25-dimensional lattice vector). A WGAN with a Wasserstein critic and an added formation-energy loss then generates new latent vectors, and the VAE decoder turns them back into crystal structures. The quantity that carries the thermodynamic argument is the formation energy $E_f=(E_{V_xO_y}-xE_V-yE_O)/(x+y)$ from Eq. (4), computed with VASP and compared to convex-hull energies; this is what separates stable, metastable, and unstable candidates. The voxel encoding plus the formation-energy constraint is what keeps the generated structures chemically valid and periodically consistent rather than arbitrary atom arrangements.

What would settle it

Recompute the two claimed below-hull V2O3 structures using the same functional, pseudopotentials, and elemental reference states used to build the Materials Project hull, with a converged k-point mesh; if their formation energies rise above the hull by more than the quoted numerical uncertainty, the central discovery claim fails. A second check is to compute phonons for these two structures in larger supercells: persistent imaginary modes would indicate dynamical instability at 0 K even if the formation energy stays below the hull.

Watch

Extended reading notes

Core claim

The central claim is that combining a WGAN with formation-energy constraints and a voxel-based VAE lets a model generate thermodynamically feasible V–O compositions that reach beyond known databases. Of the 451 generated materials (184 VO2, 152 V2O3, 115 V2O5), the paper classifies 91 as stable (negative formation energy and at most 300 meV/atom from the convex hull) and 44 as metastable (at most 500 meV/atom), a success rate the authors contrast with the looser 0.5 eV/atom cutoff used previously. The strongest specific finding is that two V2O3 structures sit below the Materials Project convex hull, implying new stable phases not previously documented; for the lowest-energy VO2 and V2O3 candidates, spin-polarized DFT+U shows a large gap in one spin channel (about 2.7 eV and 3.73 eV respectively) and metallic character in the other. Phonon calculations for selected structures show only small imaginary modes, which the paper attributes to finite supercell sizes or the known rutile-to-monoclinic transition of VO2. These results are offered as evidence that the framework accelerates discovery of functional materials, with the two below-hull V2O3 phases as the key evidence.

Load-bearing premise

The below-hull result rests on comparing formation energies computed with isolated-atom references at a k-point spacing of 0.5 inverse Angstroms and a force tolerance of 0.05 eV per Angstrom directly against the Materials Project convex hull; if that comparison carries a systematic offset of even a few tens of meV per atom, the two alleged new stable V2O3 phases could be ordinary metastable structures.

Editorial extensions

If this is right

  • If the two below-hull V2O3 phases are genuine, the known vanadium oxide phase diagram is incomplete and the generative pipeline has found stable polymorphs that conventional database screening missed.
  • A strict stability hit rate of about 20 percent suggests that latent-space generation with formation-energy constraints can produce viable candidates much more efficiently than the looser earlier benchmark.
  • The reported half-metallic VO2 and V2O3 candidates would be direct inputs for spintronics and spin-filtered electronics, provided their electronic structure survives more accurate methods.
  • The phonon results, with only minor imaginary modes, imply the generated phases are dynamically viable at practical temperatures, making at least some of them worth attempting to synthesize.
  • The same architecture can be applied to other transition-metal oxide families by changing the chemical constraints and stability criteria, extending the discovery claim beyond V-O compounds.

Reading between the lines

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

  • The below-hull claim rests entirely on aligning the DFT energies from Eq. (4) with the Materials Project's reference frame; a direct re-relaxation of the two structures with the project's own settings would settle whether they are true new phases or an artifact of the loose computational parameters.
  • The paper's 'stable' category allows up to 300 meV/atom above the hull, so most of the 91 stable candidates are not ground states; the real novelty burden is carried by the two below-hull V2O3 structures, not by the overall hit rate.
  • The half-metallicity is predicted with PBE+U at a single Hubbard value, so spin-resolved photoemission or more accurate hybrid-functional calculations would be a natural experimental or computational check before device-level claims are made.
  • A useful stress test for the framework would be to run the same pipeline on a very well-explored oxide family and count below-hull hits; a nonzero rate there would suggest the below-hull signal is systematic physics, while a zero rate would point to numerical noise in the V-O case.
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 an inverse design framework combining a Wasserstein GAN (WGAN) with a voxel-based variational autoencoder (VAE) to generate vanadium oxide compositions. The authors train the model on Materials Project data, generate 451 unique V-O structures, and classify 91 as stable and 44 as metastable under their criteria. They further claim that two V2O3 configurations have formation energies below the Materials Project convex hull, indicating previously unknown thermodynamically stable phases. Supporting analyses include PBE formation energies, spin-polarized DFT+U electronic structure calculations, and phonon dispersion calculations for selected phases. The central scientific claim is the discovery of these below-hull V2O3 phases.

Significance. If substantiated, the below-hull V2O3 claim would be a significant finding, as V2O3 is a well-studied oxide and new stable phases would be of considerable interest. The paper also provides a potentially reusable generative framework and a public GitHub repository with data and scripts, which is a strength. However, the central claim rests on a comparison between formation energies computed with isolated-atom references and a convex hull built from elemental-phase references; this mismatch, together with the loose DFT settings, means the claim is currently unsupported. The methodological novelty of the WGAN-VAE architecture is diminished by the absence of a precise definition of the 'formation energy loss' and 'stability constraints' that are said to be integrated into the WGAN.

major comments (4)
  1. [§3.4, Eq. (4)] The formation energy in Eq. (4) is defined relative to isolated atoms (E_V and E_O), whereas the Materials Project convex hull is referenced to elemental phases (bcc V and the O2 molecule). The paper never describes converting between these two reference frames, so the reported 'below the Materials Project convex hull' energies in Section 3.4 are an apples-to-oranges comparison. The difference between the two reference frames is a composition-dependent offset that can shift every structure relative to the hull. The authors must either derive and apply the conversion explicitly or recompute the hull in the same reference frame as Eq. (4). Without this, the below-hull claim is not meaningful.
  2. [§3.1, §3.4] The DFT settings used for the stability calculations (0.5 Å^-1 k-point spacing, 0.05 eV/Å force convergence, no static calculation) are too coarse to resolve energy differences of tens of meV/atom that the below-hull claim requires. No convergence tests with respect to k-point density, cutoff energy, or force threshold are provided. The Materials Project uses much denser k-point sampling and additional correction terms, so the claimed sub-hull energies may be within the numerical noise of the present calculations. The authors need to demonstrate convergence, for example by recomputing the two V2O3 candidates with denser k-point meshes and tighter force criteria, and by reporting the resulting energy changes.
  3. [§3.3] The stability criteria are not precisely defined. The text states that a structure is 'stable' when its formation energy is negative and its distance to the convex hull is ≤ 300 meV/atom, but it is unclear whether this distance is computed using the authors' own formation energies or the Materials Project hull values, and whether a negative distance (below hull) is allowed. Since the central claim concerns structures below the hull, the definition of 'distance' must be unambiguous and applied consistently. In addition, Section 3.5 refers to the V2O3 phase as having a formation energy 'close to the convex hull', which contradicts the 'below hull' wording in Section 3.4 and suggests the authors themselves are uncertain about the reference frame.
  4. [§2.5, §3.1] The paper states that the WGAN includes a 'custom formation energy loss' and 'integrated stability constraints', but neither of these components is defined mathematically or described in sufficient detail to be reproducible. Since these are presented as key innovations of the framework, the authors must provide the exact form of the loss function, how the stability constraint is enforced during training, and how the formation energy prediction is obtained and validated. Without this information, the methodological contribution cannot be assessed.
minor comments (5)
  1. [§3.5] The phonon calculations show imaginary modes, but the authors attribute them to finite-size effects without providing convergence tests. It would strengthen the paper to show how the imaginary modes change with supercell size, as is partially suggested by the supplementary information.
  2. [§3.6] The quoted band gaps (2.7 eV for VO2 and 3.73 eV for V2O3) are given to two decimal places, which overstates the precision of DFT+U band-structure calculations; consider reporting these as approximate values with a clear statement of numerical uncertainty.
  3. [§3.1] The sentence describing the phonon convergence criterion is incomplete: 'ensuring that the residual forces were below 0.01 eV·Å−1, in accordance with the convergence criterion set in our calculations' appears after a line break and should be joined into a full sentence.
  4. [§2.2] The description of the dataset construction is vague: the authors say they built on prior work [28] that curated 10,981 binary materials, but they do not specify how the substitution approach was applied to generate V-O structures or how many distinct structures were used for training. More detail is needed for reproducibility.
  5. [References] Reference [38] incorrectly cites Perdew, Burke, and Ernzerhof for the PAW method; the PAW method is from Blöchl, and the citation should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the generated structures are evaluated by independent DFT calculations, and the below-hull claim is a benchmark comparison rather than a tautology.

full rationale

The derivation chain is not circular. The WGAN-VAE is trained on Materials-Project-derived V-O structures and the WGAN includes a formation-energy loss, but the paper's stability claims are not read off from the generative model. Section 3.1 states that 'we generated 451 new materials that we use DFT for calculating their formation energies,' and Eq. (4) computes formation energies from VASP total energies; phonon and DFT+U results are likewise independent first-principles checks. The below-convex-hull V2O3 claim compares new DFT energies with the Materials Project hull. Although the training set and the hull share the Materials Project database, the generated structures are new and their energies are not fit parameters of the model, so the comparison is a benchmark rather than an identity. Concerns about the isolated-atom reference frame in Eq. (4) versus the elemental reference frame of the Materials Project hull, and about the loose k-point and force tolerances, are correctness and validation risks, not circularity: an incorrect conversion would make the claim unsupported or false, not tautological. The only self-citation, reference [24], supports a general remark about transfer learning approaches and is not load-bearing. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and no equation reduces to its own input.

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

The central claim depends on stability thresholds and reference conversions that are not fully specified. The paper uses the Materials Project convex hull as the benchmark but computes formation energies with isolated-atom references, and never states how the two are aligned. Generated structures are not independently validated beyond the authors' own DFT calculations.

free parameters (5)
  • stability threshold (stable) = 300 meV/atom hull distance
    Chosen by hand in Section 3.3 to classify generated structures as stable; more permissive than the common 0-100 meV thresholds, inflating stability counts.
  • stability threshold (metastable) = 500 meV/atom hull distance
    Chosen by hand in Section 3.3; defines metastability.
  • beta (KL weight in VAE loss)
    In Eq. (3) beta controls reconstruction vs regularization trade-off; no value or tuning procedure given.
  • Hubbard U for V 3d = 3.25 eV
    Taken from literature, not fitted in this work; affects electronic structure claims.
  • latent dimensions (site VAE 200, lattice VAE 25, WGAN 200) = 200/25/200
    Architecture hyperparameters chosen without stated optimization.
assumptions (4)
  • domain assumption PBE+U DFT with U=3.25 eV gives accurate formation energies and electronic structure for V-O phases.
    Used throughout for stability and property validation; this is an approximation, not a formal proof.
  • domain assumption The Materials Project convex hull for V-O is complete and sufficiently accurate to serve as the stability benchmark.
    The paper defines stability by distance to this hull; incomplete databases could misclassify structures.
  • domain assumption The voxel representation with 64x64x64 site grids and 32x32x32 lattice grid fully captures the degrees of freedom needed to reconstruct valid crystals.
    If the representation loses periodicity or symmetry, generated structures may be unphysical; the paper provides no reconstruction validation metrics.
  • ad hoc to paper Negative formation energy relative to isolated atoms is a meaningful stability criterion.
    Eq. (4) uses isolated atom references, but standard stability uses elemental references; no conversion is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accelerated Discovery of Vanadium Oxide Compositions: A WGAN-VAE Framework for Materials Design." pith.science (2026). https://pith.science/paper/R4YL5H7G

@misc{pith2026250104604,
  author       = {Pith},
  title        = {Pith review of: Accelerated Discovery of Vanadium Oxide Compositions: A WGAN-VAE Framework for Materials Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R4YL5H7G}},
  note         = {Machine review of arXiv:2501.04604}
}
read the original abstract

The discovery of novel materials with tailored electronic properties is crucial for modern device technologies, but time-consuming empirical methods hamper progress. We present an inverse design framework combining an enhanced Wasserstein Generative Adversarial Network (WGAN) with a specialized Variational Autoencoder (VAE) to accelerate the discovery of stable vanadium oxide (V-O) compositions. Our approach features (1) a WGAN with integrated stability constraints and formation energy predictions, enabling direct generation of thermodynamically feasible structures, and (2) a refined VAE capturing atomic positions and lattice parameters while maintaining chemical validity. Applying this framework, we generated 451 unique V-O compositions, with 91 stable and 44 metastable under rigorous thermodynamic criteria. Notably, we uncovered several novel V2O3 configurations with formation energies below the Materials Project convex hull, revealing previously unknown stable phases. Detailed spin-polarized DFT+U calculations showed distinct electronic behaviors, including promising half-metallic characteristics. Our approach outperforms existing methods in both quality and stability, demonstrating about a 20 percent stability rate under strict criteria compared to earlier benchmarks. Additionally, phonon calculations performed on selected compositions confirm dynamic stability: minor imaginary modes at 0 K likely stem from finite-size effects or known phase transitions, suggesting that these materials remain stable or metastable in practical conditions. These findings establish our framework as a powerful tool for accelerated materials discovery and highlight promising V-O candidates for next-generation electronic devices.

Figures

Figures reproduced from arXiv: 2501.04604 by the authors.

Figure 1
Figure 1. Research trends and applications of vanadium oxide (V [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 13
Figure 13. Spin-polarized DFT+U band structure (left) and spin-resolved DOS (right) for a triclinic VO2 phase. states show a distinct gap around the Fermi level while spin-up states maintain finite den￾sity. Such electronic structure characteristics strongly suggest half-metallic behavior, where one spin channel provides conduction pathways while the other remains insulating. The distinct spin-dependent bandgaps make this mate… view at source ↗
Figure 14
Figure 14. The triclinic V2O3 phase is represented with the spin-polarized DFT+U band structure and the spin-resolved DOS. term. The DOS calculations indicate the presence of accessible states at and around the Fermi level for both spin orientations, suggesting that electrons in both spin channels can participate in conduction processes. This electronic structure does not support half-metallic [PITH_FULL_IMAGE:figures/full_fi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 24 canonical work pages

  1. [1]

    Elsie S Place, Nicholas D Evans, Molly M Stevens, Complexity in biomaterials for tissue engineering, Nat. Mater. 8. 6 (2009) 457-470. https://doi.org/10.1038/nmat2441

    work page doi:10.1038/nmat2441 2009

  2. [2]

    https://doi.org/10.1016/j.mtelec.2023.100052

    Muhammad Naqi, Yongin Cho, Arindam Bala, Sunkook Kim, The trend of synthesized 2D materials toward artificial intelligence: Memory technology and neuromorphic computing , Ma- terials Today Electronics (2023) 100052. https://doi.org/10.1016/j.mtelec.2023.100052

    arXiv 2023

  3. [3]

    Peng Hu, Ping Hu, Tuan Duc Vu, Ming Li, Shancheng Wang, Yujie Ke, Xianting Zeng, Liqiang Mai, Yi Long, Vanadium oxide: phase diagrams, structures, synthesis, and applications, Chem. Rev. 123. 8 (2023) 4353-4415. https://doi.org/10.1021/acs.chemrev.2c00546

    work page doi:10.1021/acs.chemrev.2c00546 2023

  4. [4]

    Benjamin Sanchez-Lengeling, Al ´an Aspuru-Guzik, Inverse molecular design using ma - chine learning: Generative models for matter engineering, Sci. 361. 6400 (2018) 360-365. https://doi.org/10.1126/science.aat2663

    doi:10.1126/science.aat2663 2018

  5. [5]

    Daniel Wines, Kamal Choudhary, Adam J Biacchi, Kevin F Garrity, Francesca Tavazza, High- throughput DFT-based discovery of next generation two-dimensional (2D) superconductors, Nano Lett. 23. 3 (2023) 969-978. https://doi.org/10.1021/acs.nanolett.2c04420

    work page doi:10.1021/acs.nanolett.2c04420 2023

  6. [6]

    Guanjian Cheng, Xin -Gao Gong, Wan -Jian Yin, Crystal structure prediction by com - bining graph network and optimization algorithm , Nat. Commun. 13. 1 (2022) 1492. https://doi.org/10.1038/s41467-022-29241-4

    work page doi:10.1038/s41467-022-29241-4 2022

  7. [7]

    Yanchao Wang, Jian Lv, Li Zhu, Yanming Ma, CALYPSO: A method for crys- tal structure prediction , Comput. Phys. Commun. 183. 10 (2012) 2063 -2070. https://doi.org/10.1016/j.cpc.2012.05.008

    doi:10.1016/j.cpc.2012.05.008 2012

  8. [8]

    Aryan Deshwal, Cory M Simon, Janardhan Rao Doppa, Bayesian optimization of nanoporous materials, Mol. Syst. Des. Eng. 6. 12 (2021) 1066-1088. https://doi.org/10.1039/D1ME00093D

    work page doi:10.1039/d1me00093d 2021

Show all 50 references
  1. [9]

    Tomoki Yamashita, Nobuya Sato, Hiori Kino1, Takashi Miyake, Koji Tsuda1, Tamio Oguchi, Crystal structure prediction accelerated by Bayesian optimization , Phys. Rev. Mater. 2. 1 (2018) 013803. https://doi.org/10.1103/PhysRevMaterials.2.013803

    2018 doi

  2. [10]

    Andrew J Mannix, Xiang-Feng Zhou, Brian Kiraly, Joshua D Wood, Diego Alducin, Benjamin D Myers, Xiaolong Liu, Brandon L Fisher, Ulises Santiago, Jeffrey R Guest, Miguel Jose Yacaman, Arturo Ponce, Artem R Oganov, Mark C Hersam, Nathan P Guisinger, Synthesis of borophenes: Anis...

    2015

  3. [11]

    Xiangyang Liu, Haiyang Niu, Artem R Oganov, COPEX: co -evolutionary crystal struc - ture prediction algorithm for complex systems , Npj Comput. Mater. 7. 1 (2021) 199. https://doi.org/10.1038/s41524-021-00668-5

    2021 doi

  4. [12]

    Artem R Oganov, Colin W Glass, Crystal structure prediction using ab initio evo - lutionary techniques: Principles and applications , J. Chem. Phys. 124. 24 (2006). https://doi.org/10.1063/1.2210932

    2006 doi

  5. [13]

    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 innovation...

    2013 doi

  6. [14]

    Scott Kirklin, James E Saal, Bryce Meredig, Alex Thompson, Jeff W Doak, Muratahan Aykol, Stephan Ru¨hl, Chris Wolverton, The Open Quantum Materials Database (OQMD): assessing the accuracy of DFT formation energies, Npj Comput. Mater. 1. 1 (2015) 1 -15. https://doi.org/10.1038/...

    2015 doi

  7. [15]

    Yann LeCun, Yoshua Bengio, Geoffrey Hinton, Deep learning, Nat. 521. 7553 (2015) 436-444. https://doi.org/10.1038/nature14539

    2015 doi

  8. [16]

    https://doi.org/10.1016/j.mtelec.2023.100027

    Zhiheng Cheng, Xingran Xie, Yimin Yang, Chaolun Wang, Chen Luo, Hengchang Bi, Yan Wang, Junhao Chu, Xing Wu, Neural network approach for ferroelectric hafnium oxide phase identification at the atomistic scale, Materials Today Electronics 3 (2023) 100027. https://doi.org/10.101...

    2023

  9. [17]

    https://doi.org/10.48550/arXiv.1312.6114

    Diederik P Kingma, Max Welling, Auto-encoding variational bayes , arXiv preprint arXiv:1312.6114 (2013). https://doi.org/10.48550/arXiv.1312.6114

    2013 doi arXiv

  10. [18]

    Neural Inf

    Ian Goodfellow, Jean Pouget -Abadie, Mehdi Mirza, Bing Xu, David Warde -Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, Generative adversarial nets, Adv. Neural Inf. Process. Syst. (NIPS). 27 (2014)

    2014

  11. [19]

    https://doi.org/10.48550/arXiv.2411.09429

    Xiao-Qi Han, Xin-De Wang, Meng-Yuan Xu, Zhen Feng, Bo-Wen Yao, Peng-Jie Guo, Ze-Feng Gao, Zhong-Yi Lu, AI-driven inverse design of materials: Past, present and future, arXiv preprint arXiv:2411.09429 (2024). https://doi.org/10.48550/arXiv.2411.09429

    2024 doi arXiv

  12. [20]

    Paul Raccuglia, Katherine C Elbert, Philip D F Adler, Casey Falk, Malia B Wenny, Aurelio Mollo, Matthias Zeller, Sorelle A Friedler, Joshua Schrier, Alexander J Norquist, Machine- learning-assisted materials discovery using failed experiments, Nat. 533. 7601 (2016) 73-76. http...

    2016 doi

  13. [21]

    https://doi.org/10.48550/arXiv.2210.11931

    Elton Pan, Christopher Karpovich, Elsa Olivetti, Deep reinforcement learning for inverse inorganic materials design , arXiv preprint arXiv:2210.11931 (2022). https://doi.org/10.48550/arXiv.2210.11931

    2022 doi arXiv

  14. [22]

    Victor Fung, Jiaxin Zhang, Guoxiang Hu, P Ganesh, Bobby G Sumpter, Inverse design of two-dimensional materials with invertible neural networks, Npj Comput. Mater. 7. 1 (2021)

    2021

  15. [23]

    Neural Inf

    Chongxuan LI, Taufik Xu, Jun Zhu, Bo Zhang, Triple generative adversarial nets, Adv. Neural Inf. Process. Syst. (NIPS). 30 (2017)

    2017

  16. [24]

    Danial Ebrahimzadeh, Sarah Sharif, Yaser Banad, How transfer learning is used in generative models for image classification: improved accuracy, Signal Image Video P. 19. 1 (2025) 1-9. https://doi.org/10.1007/s11760-024-03673-5

    2025 doi

  17. [25]

    Karen E Willcox, Omar Ghattas, Patrick Heimbach, The imperative of physics-based mod- eling and inverse theory in computational science, Nat. Comput. Sci. 1. 3 (2021) 166-168. https://doi.org/10.1038/s43588-021-00040-z

    2021 doi

  18. [26]

    Shuaihua Lu, Qionghua Zhou, Xinyu Chen, Zhilong Song, Jinlan Wang, Inverse design with deep generative models: next step in materials discovery, Natl. Sci. Rev. 9. 8 (2022) nwac111. https://doi.org/10.1093/nsr/nwac111

    2022 doi

  19. [27]

    Rafael G´omez -Bombarelli, Jennifer N Wei, David Duvenaud, Jo s´e Miguel Hern´andez -Lobato, Benjam´ın S´anchez -Lengeling, Dennis Sheberla, OrcidJorge Aguilera -Iparraguirre, Timo- thy D Hirzel, Ryan P Adams, Al ´an Aspuru-Guzik, Automatic chemical design using a data-driven ...

    2018 doi

  20. [28]

    Juhwan Noh, Jaehoon Kim, Helge S Stein,Benjamin Sanchez -Lengeling, John M Gregoire, Alan Aspuru-Guzik, Yousung Jung, Inverse design of solid-state materials via a continuous representation, matt. 1. 5 (2019) 1370-1384. https://doi.org/10.1016/j.matt.2019.08.017

    2019 doi

  21. [29]

    Frank No´e, Simon Olsson, Jonas K¨ohler, Hao Wu, Boltzmann generators: Sampling equi- librium states of many-body systems with deep learning, Sci. 365. 6457 (2019) eaaw1147. https://doi.org/10.1126/science.aaw1147

    2019 doi

  22. [30]

    https://doi.org/10.48550/arXiv.1810.11203

    Asma Nouira, Nataliya Sokolovska, Jean -Claude Crivello, Crystalgan: learning to dis - cover crystallographic structures with generative adversarial networks, arXiv preprint arXiv:1810.11203 (2018). https://doi.org/10.48550/arXiv.1810.11203

    2018 doi arXiv

  23. [31]

    Teng Long, Nuno M Fortunato, Ingo Opahle, Yixuan Zhang, Ilias Samathrakis, Chen Shen, Oliver Gutfleisch, Hongbin Zhang, Constrained crystals deep convolutional generative adver- sarial network for the inverse design of crystal structures, Npj Comput. Mater. 7. 1 (2021) 66. htt...

    2021 doi

  24. [32]

    David Weininger, SMILES, a chemical language and information system. 1. Introduc - tion to methodology and encoding rules , J. Chem. Inf. Comput. 28. 1 (1988) 31 -36. https://doi.org/10.1021/ci00057a005

    1988 doi

  25. [33]

    Kristof T Schutt, Henning Glawe, Felix Brockherde, Antonio Sanna, Klaus -Robert Muller, Eberhard KU Gross, How to represent crystal structures for machine learning: To- wards fast prediction of electronic properties , Phys. Rev. B. 89. 20 (2014) 205118. https://doi.org/10.1103...

    2014 doi

  26. [34]

    Neural Inf

    Yann LeCun, Bernhard E Boser, John S Denker, Donnie Henderson, R E Howard, Wayne E Hubbard, Lawrence D Jackel, Handwritten Digit Recognition with a Back-Propagation Net- work, Adv. Neural Inf. Process. Syst. (NIPS). 2 (1989) 23

    1989

  27. [35]

    (2016) 770-778

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, Deep residual learning for image recognition, Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR). (2016) 770-778

    2016

  28. [36]

    (2017) 214-223

    Martin Arjovsky, Soumith Chintala, L´eo n Bottou, Wasserstein generative adversarial net- works, Proceedings of the 34th International Conference on Machine Learning (PMLR). (2017) 214-223

    2017

  29. [37]

    John P Perdew, Kieron Burke, Matthias Ernzerhof, Generalized gradi- ent approximation made simple , Phys. Rev. Lett. 77. 18 (1996) 3865 -3868. https://doi.org/10.1103/PhysRevLett.77.3865

    1996 doi

  30. [38]

    John P Perdew, Kieron Burke, Matthias Ernzerhof, Projector augmented-wave method, Phys. Rev. B. 50. 24 (1994) 17953-17979. https://doi.org/10.1103/PhysRevB.50.17953

    1994 doi

  31. [39]

    Georg Kresse, J Furthm u¨ller, Software VASP, Vienna (1999), Phys. Rev. B. 54. 11 (1996) 169

    1999

  32. [40]

    Vei Wang, Nan Xu, Jin-Cheng Liu, Gang Tang, Wen-Tong Geng, VASPKIT: A user-friendly interface facilitating high -throughput computing and analysis using VASP code , Comput. Phys. Commun. 267 (2021) 108033. https://doi.org/10.1016/j.cpc.2021.108033

    2021

  33. [41]

    Phys.: Condens

    Atsushi Togo, Laurent Chaput, Terumasa Tadano, Isao TanakaVei, Implementation strate- gies in phonopy and phono3py , J. Phys.: Condens. Matter. 35. 35. (2023) 353001. https://dx.doi.org/10.1088/1361-648X/acd831

    2023 doi

  34. [42]

    Atsushi Togo, First-principles Phonon Calculations with Phonopy and Phono3py , J. Phys. Soc. Jpn. 92. 1. (2023) 012001. https://doi.org/10.7566/JPSJ.92.012001

    2023 doi

  35. [43]

    https://doi.org/10.48550/arXiv.2411.09429 https://doi.org/10.48550/arXiv.1910.00927

    Maciej Wiatrak, Stefano V Albrecht, Andrew Nystrom, Stabilizing Genera- tive Adversarial Networks: A Survey , arXiv preprint arXiv:1910.00927v2 (2019). https://doi.org/10.48550/arXiv.2411.09429 https://doi.org/10.48550/arXiv.1910.00927

    2019 doi arXiv

  36. [44]

    Mumtaz M Qazilbash, Markus Brehm, Byung-Gyu Chae, P-C Ho, Gregory O Andreev, Bong- Jun Kim, Sun Jin Yun, A V Balatsky, M B Maple, F Keilmann, H T Kim, Mott Transition in VO2 Revealed by Infrared Spectroscopy and Nano -Imaging, Science. 318. 5857 (2007) 1750-

    2007

  37. [45]

    F J Morin, Oxides which show a metal -to-insulator transition at the Neel temperature , Physical Review Letters. 3. 1 (1959) 34. https://doi.org/10.1103/PhysRevLett.3.34

    1959 doi

  38. [46]

    Zhi W Hua, Tan Zheng, Jing X Sang, Xiao F Xu, Bin H Wu, Lian F Wei, Impact of anharmonic effects on two phases of VO2 thin-film phase transition materials, J. Phys. D: Appl. Phys. 54. 35 (2021) 355302. https://doi.org/10.1088/1361-6463/ac02f7

    2021 doi

  39. [47]

    Jason M Munro, Katherine Latimer, Matthew K Horton, Shyam Dwaraknath, Kristin A Pers- son, An improved symmetry-based approach to reciprocal space path selection in band structure calculations, npj Comput. Mater. 6. 1 (2020) 112. https://doi.org/10.1038/s41524-020-00383-7

    2020 doi

  40. [200]

    https://doi.org/10.1038/s41524-021-00670-x. 22

    doi

  41. [1516]

    https://doi.org/10.1126/science.aad1080. 21

    doi

  42. [1753]

    https://doi.org/10.1126/science.11501

    doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.