arxiv: 2605.01286 · v1 · submitted 2026-05-02 · ❄️ cond-mat.mtrl-sci

Recognition: unknown

Inverse Materials Design via Joint Generation of Crystal Structures and Local Electronic Descriptors

Ibuki Okuda , Izumi Takahara , Teruyasu Mizoguchi

Authors on Pith no claims yet

Pith reviewed 2026-05-09 15:01 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords inverse designcrystal structure generationdiffusion modelsBader chargeatomic density of statesconditional generationmaterials discovery

0 comments

The pith

Jointly denoising crystal structures and local electronic descriptors in a diffusion model improves success rates for generating materials with targeted band gaps or formation energies.

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

The paper establishes that a diffusion-based generative model can simultaneously produce crystal structures and site-specific electronic descriptors such as Bader charges and atomic density of states. This joint process raises the fraction of outputs that meet specified property targets compared with structure-only baselines. It also increases the share of results that satisfy uniqueness, novelty, thermodynamic stability, and physical validity. A control experiment with dummy variables shows the gains trace to the electronic information itself rather than merely adding extra per-site inputs. The approach therefore reduces the usual tension between enforcing a target property and keeping generated crystals realistic.

Core claim

By training a shared score network to denoise both structural coordinates and local electronic descriptors (Bader charge plus atomic DOS) in one trajectory, the model generates crystals that satisfy band-gap or formation-energy conditions at higher rates than structure-only baselines while also raising the fraction of outputs that meet the VSUN criteria of uniqueness, novelty, stability, and validity. The generated Bader charges match DFT references with 0.055 e mean absolute error on stable structures, and the generated atomic DOS reproduces the main spectral shape of the reference data, although finer features vary by element.

What carries the argument

A shared score network inside a conditional diffusion process that jointly denoises crystal lattice and atomic-position variables together with site-resolved Bader charges and atomic density of states.

If this is right

Conditioned generation succeeds more often when electronic descriptors guide the structural trajectory.
The fraction of outputs meeting uniqueness, novelty, stability, and validity rises under the joint scheme.
The electronic content of the descriptors, not merely their presence as extra variables, accounts for the measured gains.
Generated Bader charges on stable structures align with DFT calculations to within 0.055 e mean absolute error.

Where Pith is reading between the lines

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

The same joint framework could be tested with additional local descriptors such as magnetic moments to target spin-dependent properties.
If the generated atomic DOS proves sufficiently accurate, it could serve as an initial electronic-structure estimate that reduces the number of full DFT relaxations needed in screening workflows.
Extending the conditioning to other global properties such as dielectric constants might further demonstrate whether local electronic guidance generalizes beyond band gap and formation energy.

Load-bearing premise

The chosen local electronic descriptors supply guidance that steers denoising toward electronically plausible structures without introducing new inconsistencies or selection biases beyond those already tested by the dummy-variable control.

What would settle it

Running the same conditioned generation tasks on a fresh set of target band-gap and formation-energy values or on a different elemental composition space and finding no improvement in success rate or VSUN fraction over the structure-only baseline would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.01286 by Ibuki Okuda, Izumi Takahara, Teruyasu Mizoguchi.

**Figure 1.** Figure 1: Schematic illustration of the proposed diffusion framework for the joint generation of crystal structures view at source ↗

**Figure 2.** Figure 2: Band-gap and formation-energy distributions and conditional generation success rates for VSUN struc view at source ↗

**Figure 3.** Figure 3: Novelty of VSUN structures generated by the view at source ↗

**Figure 4.** Figure 4: Comparison between generated and DFTcalculated Bader charges for atoms in the generated structures. a. Overall correlation for all atoms across all generated structures. Atoms belonging to SUN structures (MAE: 5.5 × 10−2 e) are shown using a density colormap, and those from non-SUN structures are plotted as gray points. b. Element-resolved correlations for O, S, Sr, and Au. In each panel, atoms from SUN… view at source ↗

**Figure 5.** Figure 5: Evaluation of atomic DOS generation quality using the Wasserstein distance view at source ↗

**Figure 6.** Figure 6: Compact schematic of the convolutional autoencoder used for atomic-DOS compression. The model view at source ↗

**Figure 7.** Figure 7: Element-resolved generation accuracy visualized on the periodic table. view at source ↗

read the original abstract

Inverse design of inorganic crystals, in which structures are generated to satisfy a target property while preserving diversity and physical plausibility, remains more demanding than ab initio generation, as property conditioning can degrade the structural quality that current generative models otherwise achieve. We propose a diffusion framework that jointly denoises crystal-structure variables and site-resolved local electronic descriptors through a shared score network. As representative descriptors, we adopt Bader charge and atomic density of states (atomic DOS). Under both band-gap and formation energy conditioned generation, the joint models achieved higher success rates than the structure-only baseline in most target conditions, while simultaneously increasing the fraction of generated structures that satisfy uniqueness, novelty, thermodynamic stability, and physical validity (VSUN criteria). A dummy-variable control confirms that these gains originate from the electronic content of the descriptors rather than from auxiliary site-wise variables. The generated Bader charges agree with DFT references with an MAE of 5.5e-2 e on stable structures, and the generated atomic DOS captures the coarse spectral profile of the DFT reference around the modal accuracy range, although finer details and accuracy vary with elemental species. These results establish local electronic descriptors as effective generative variables that serve two complementary roles: broadening the explored materials space through increased structural diversity, and mitigating the trade-off between property targeting and structural quality by guiding the structural trajectory toward electronically plausible configurations during joint denoising.

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.

Desk Editor's Note private letter to a colleague

Joint denoising of crystal structures with Bader charges and atomic DOS yields measurable gains in success rate and VSUN metrics over a structure-only baseline, but the dummy-variable control leaves room for doubt on whether the electronic content is truly doing the steering.

read the letter

The paper's core result is that a diffusion model jointly denoising atomic positions, lattice, and site-wise electronic descriptors (Bader charge plus atomic DOS) produces higher success rates under band-gap and formation-energy conditioning than a structure-only version, while also lifting the fraction of outputs that meet uniqueness, novelty, stability, and validity criteria. A dummy-variable ablation is offered as evidence that the improvement comes from the actual descriptor content rather than the mere presence of extra per-site channels. They also report an MAE of 0.055 e for the generated Bader charges against DFT on stable structures and note that the atomic DOS reproduces the broad spectral shape, though finer features vary by element. That combination of joint generation plus a control experiment is the clearest technical step beyond prior structure-only diffusion work for crystals. The numbers on success and VSUN are concrete enough to be worth checking, and the choice of local electronic descriptors as generative variables is a reasonable attempt to inject electronic plausibility directly into the denoising path. The dummy test is a positive move toward isolating the source of the gains. The main weaknesses sit in the methods and controls. The abstract gives no information on training-test splits, hyperparameter search, or whether the reported VSUN improvements clear a significance threshold. The DOS comparison stays qualitative, which limits how much weight one can place on it. More critically, the dummy control may not be matched on variance or per-element distributions; if the dummies are simpler than the real descriptors, the extra statistical structure in the real variables could produce the observed lift without the descriptors actually guiding the model toward electronically consistent configurations. That possibility is not ruled out by what is shown. The work is aimed at groups already running diffusion or generative models for inorganic crystals and looking for ways to tighten property conditioning without sacrificing diversity. Readers who care about practical inverse-design pipelines will find the setup and the control experiment worth examining. It is coherent enough on its own terms and shows honest engagement with the quality-diversity trade-off, so it deserves a serious referee rather than a desk reject, though any review will need to press for fuller methods disclosure and tighter ablations on the dummy variables.

Referee Report

3 major / 3 minor

Summary. The paper proposes a diffusion framework for inverse crystal design that jointly denoises atomic positions, lattice parameters, and site-resolved local electronic descriptors (Bader charges and atomic DOS) via a shared score network. Conditioned on target band gaps or formation energies, the joint models report higher success rates than a structure-only baseline in most cases, along with improved fractions of structures satisfying uniqueness, novelty, thermodynamic stability, and physical validity (VSUN). A dummy-variable control is used to attribute gains to electronic content rather than auxiliary variables. Generated Bader charges achieve MAE of 5.5e-2 e versus DFT on stable structures, while atomic DOS captures coarse spectral features.

Significance. If the central claims hold after addressing controls and reporting gaps, the work would be significant for materials generation: it shows that local electronic descriptors can simultaneously improve property targeting and structural quality metrics, addressing a known trade-off in conditioned diffusion models. The explicit baseline comparison and dummy control are strengths that allow direct assessment of the electronic contribution.

major comments (3)

[Methods / Experiments] The dummy-variable control is load-bearing for the claim that gains originate from electronic content (abstract and results section). The manuscript does not specify whether dummies are sampled from the empirical per-element marginals and variances of real Bader charges/DOS or implemented as simpler random/constant auxiliaries; mismatched statistics could produce the observed VSUN and success-rate improvements without true electronic guidance.
[Results] Quantitative claims of improved success rates and VSUN fractions lack details on training/validation splits, hyperparameter choices, number of independent runs, and statistical significance testing. These omissions make it difficult to evaluate whether the reported gains over the structure-only baseline are robust or could arise from optimization variance.
[Results / Descriptor Accuracy] Atomic DOS accuracy is described only qualitatively as capturing the 'coarse spectral profile' within 'modal accuracy range.' Providing element-specific quantitative metrics (e.g., MAE on integrated DOS or peak-position errors) would be needed to support the assertion that generated descriptors are physically meaningful and non-redundant with structure.

minor comments (3)

[Methods] Notation for the joint score network and conditioning mechanism could be made more explicit (e.g., clarifying how property targets are injected alongside the electronic descriptors).
[Figures] Figure captions should include axis labels, error bars if present, and explicit definitions of VSUN criteria for each panel.
[Introduction / Related Work] A few citations to recent crystal diffusion baselines appear incomplete; ensure all compared methods are referenced with their arXiv or publication details.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and insightful report on our manuscript. We have addressed each major comment point by point below, indicating where revisions will be made to the manuscript.

read point-by-point responses

Referee: [Methods / Experiments] The dummy-variable control is load-bearing for the claim that gains originate from electronic content (abstract and results section). The manuscript does not specify whether dummies are sampled from the empirical per-element marginals and variances of real Bader charges/DOS or implemented as simpler random/constant auxiliaries; mismatched statistics could produce the observed VSUN and success-rate improvements without true electronic guidance.

Authors: We agree with the referee that the details of the dummy-variable control are critical to the interpretation of our results. The manuscript currently lacks an explicit description of the dummy construction procedure. In the revised manuscript, we will add this information to the Methods section, specifying that the dummy variables are sampled independently for each site from the empirical per-element marginal distributions (including means and variances) of the Bader charges and atomic DOS computed on the training data. This design ensures the dummies have matching marginal statistics to the real descriptors but without the joint structural-electronic correlations present in the actual data. We believe this addition will fully address the concern and strengthen the control experiment. revision: yes
Referee: [Results] Quantitative claims of improved success rates and VSUN fractions lack details on training/validation splits, hyperparameter choices, number of independent runs, and statistical significance testing. These omissions make it difficult to evaluate whether the reported gains over the structure-only baseline are robust or could arise from optimization variance.

Authors: We acknowledge that the current manuscript does not provide sufficient details on the experimental protocol to allow full assessment of robustness. We will revise the manuscript to include explicit information on the training and validation data splits used, the specific hyperparameter values selected for the model and training process, the number of independent runs performed with different random seeds, and the statistical tests (such as t-tests) applied to evaluate the significance of the differences in success rates and VSUN metrics between the joint and baseline models. These revisions will be placed in the Results and Methods sections as appropriate. revision: yes
Referee: [Results / Descriptor Accuracy] Atomic DOS accuracy is described only qualitatively as capturing the 'coarse spectral profile' within 'modal accuracy range.' Providing element-specific quantitative metrics (e.g., MAE on integrated DOS or peak-position errors) would be needed to support the assertion that generated descriptors are physically meaningful and non-redundant with structure.

Authors: We agree that additional quantitative metrics would enhance the evaluation of the generated atomic DOS. The manuscript focuses on the coarse features because they are the most relevant for the generative guidance, but we recognize the value of element-specific analysis. In the revised manuscript, we will include quantitative metrics for the atomic DOS, such as mean absolute errors on the integrated DOS and peak position errors, reported on a per-element basis for the most common elements in the dataset. This will be added to the Results section along with the existing qualitative description. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results rest on explicit baselines and controls, not self-referential derivations.

full rationale

The paper's central claims consist of empirical performance metrics (success rates, VSUN fractions) obtained by comparing a joint structure+descriptor diffusion model against an explicit structure-only baseline and a dummy-variable control. These comparisons are external to the model's internal variables and do not reduce any prediction or uniqueness statement to a fitted parameter or self-citation by construction. No equations, ansatzes, or uniqueness theorems are presented that loop back to the paper's own inputs; the dummy control is described as an independent verification that the observed gains arise from electronic content rather than auxiliary variables. The derivation chain is therefore self-contained against the reported benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard assumptions of diffusion models and DFT as ground truth, with no free parameters, new entities, or ad-hoc axioms introduced in the abstract.

axioms (2)

domain assumption Diffusion models can be extended to joint multi-variable denoising of structure and electronic descriptors via a shared score network.
Invoked in the proposal of the joint framework.
domain assumption Bader charge and atomic DOS are representative and sufficient local electronic descriptors for guiding structural generation.
Stated as representative descriptors adopted in the work.

pith-pipeline@v0.9.0 · 5552 in / 1377 out tokens · 48690 ms · 2026-05-09T15:01:03.169790+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

49 extracted references · 5 canonical work pages · 1 internal anchor

[1]

Inverse molecular design using machine learning: Generative models for matter engineer- ing.Science, 361(6400):360–365, 2018

Benjamin Sanchez-Lengeling and Alán Aspuru- Guzik. Inverse molecular design using machine learning: Generative models for matter engineer- ing.Science, 361(6400):360–365, 2018

2018
[2]

Inverse design with deep generative models: next step in materials discovery

Shuaihua Lu, Qionghua Zhou, Xinyu Chen, Zhilong Song, and Jinlan Wang. Inverse design with deep generative models: next step in materials discovery. National Science Review, 9(8):nwac111, 2022

2022
[3]

Pierre-Paul De Breuck, Hai-Chen Wang, Gian- Marco Rignanese, Silvana Botti, and Miguel A. L. Marques. Generative ai for crystal structures: a review.npj Computational Materials, 11:370, 2025

2025
[4]

Schoen- holz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk

Amil Merchant, Simon Batzner, Samuel S. Schoen- holz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for ma- terials discovery.Nature, 624:80–85, 2023

2023
[5]

Stefano Curtarolo, Wahyu Setyawan, Gus L. W. Hart, Michal Jahnátek, Roman V. Chepulskii, Richard H. Taylor, Shidong Wang, Junkai Xue, Kesong Yang, Ohad Levy, Michael J. Mehl, Harold T. Stokes, Denis O. Demchenko, and Dane Morgan. AFLOW: An automatic framework for high-throughput materials discovery.Computa- tional Materials Science, 58:218–226, 2012

2012
[6]

Stein, Ben- jamin Sanchez-Lengeling, John M

Juhwan Noh, Jaehoon Kim, Helge S. Stein, Ben- jamin Sanchez-Lengeling, John M. Gregoire, Alán Aspuru-Guzik, and Yousung Jung. Inverse design of solid-state materials via a continuous represen- tation.Matter, 1(5):1370–1384, 2019

2019
[7]

Court, Batuhan Yildirim, Apoorv Jain, andJacquelineM.Cole

Callum J. Court, Batuhan Yildirim, Apoorv Jain, andJacquelineM.Cole. 3-Dinorganiccrystalstruc- ture generation and property prediction via repre- sentation learning.Journal of Chemical Informa- tion and Modeling, 60(10):4518–4535, 2020

2020
[8]

Aberle, Shijing Sun, Xiaonan Wang, Yi Liu, Qianxiao Li, Senthilnath Jayavelu, Kedar Hippalgaonkar, Yousung Jung, and Tonio Buonassisi

Zekun Ren, Siyu Isaac Parker Tian, Juhwan Noh, Felipe Oviedo, Guangzong Xing, Jiali Li, Qiaohao Liang, Ruiming Zhu, Armin G. Aberle, Shijing Sun, Xiaonan Wang, Yi Liu, Qianxiao Li, Senthilnath Jayavelu, Kedar Hippalgaonkar, Yousung Jung, and Tonio Buonassisi. An invertible crystallo- graphic representation for general inverse design of inorganic crystals ...

2022
[9]

CrystalGAN: Learning to discover crystallographic structures with generative adver- sarial networks.arXiv preprint arXiv:1810.11203, 2018

Asma Nouira, Nataliya Sokolovska, and Jean- Claude Crivello. CrystalGAN: Learning to discover crystallographic structures with generative adver- sarial networks.arXiv preprint arXiv:1810.11203, 2018

work page arXiv 2018
[10]

Generative ad- versarial networks for crystal structure prediction

Sungwon Kim, Juhwan Noh, Geun Ho Gu, Alán Aspuru-Guzik, and Yousung Jung. Generative ad- versarial networks for crystal structure prediction. ACS Central Science, 6(8):1412–1420, 2020

2020
[11]

Benjamin Kurt Miller, Ricky T. Q. Chen, Anuroop Sriram, and Brandon M. Wood. FlowMM: Gen- erating materials with Riemannian flow matching. InForty-first International Conference on Machine Learning, 2024

2024
[12]

Generative inverse design of crystal struc- tures via diffusion models with transformers.arXiv preprint arXiv:2406.09263, 2024

Izumi Takahara, Kiyou Shibata, and Teruyasu Mi- zoguchi. Generative inverse design of crystal struc- tures via diffusion models with transformers.arXiv preprint arXiv:2406.09263, 2024

work page arXiv 2024
[13]

Deep learning generative model for crystal structure prediction.npj Computational Materials, 10:254, 2024

Xiaoshan Luo, Zhenyu Wang, Pengyue Gao, Jian Lv, Yanchao Wang, Changfeng Chen, and Yanming Ma. Deep learning generative model for crystal structure prediction.npj Computational Materials, 10:254, 2024

2024
[14]

Crystal dif- fusion variational autoencoder for periodic material generation

Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola. Crystal dif- fusion variational autoencoder for periodic material generation. InInternational Conference on Learn- ing Representations, 2022

2022
[15]

Crystal struc- turepredictionbyjointequivariantdiffusion

Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. Crystal struc- turepredictionbyjointequivariantdiffusion. InAd- vances in Neural Information Processing Systems, volume 36, 2023

2023
[16]

Space group constrained crystal gener- ation

Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. Space group constrained crystal gener- ation. InThe Twelfth International Conference on Learning Representations, 2024

2024
[17]

SymmCD: Symmetry-preserving crystal generation with diffusion models

Daniel Levy, Siba Smarak Panigrahi, Sékou-Oumar Kaba, Qiang Zhu, Kin Long Kelvin Lee, Mikhail Galkin, Santiago Miret, and Siamak Ravanbakhsh. SymmCD: Symmetry-preserving crystal generation with diffusion models. InThe Thirteenth Inter- national Conference on Learning Representations, 2025

2025
[18]

A gener- ative model for inorganic materials design.Nature, 639:624–632, 2025

Claudio Zeni, Robert Pinsler, Daniel Zügner, An- drew Fowler, Matthew Horton, Xiang Fu, Zilong 11 Wang, Aliaksandra Shysheya, Jonathan Crabbé, Shoko Ueda, Roberto Sordillo, Lixin Sun, Jake Smith, Bichlien Nguyen, Hannes Schulz, Sarah Lewis, Chin-Wei Huang, Ziheng Lu, Yichi Zhou, Han Yang, Hongxia Hao, Jielan Li, Chunlei Yang, Wenjie Li, Ryota Tomioka, and...

2025
[19]

Gleason, Ali Ram- laoui, Andy Xu, Georgia Channing, Daniel Levy, Clémentine Fourrier, Nikita Kazeev, Chaitanya K

Siddharth Betala, Samuel P. Gleason, Ali Ram- laoui, Andy Xu, Georgia Channing, Daniel Levy, Clémentine Fourrier, Nikita Kazeev, Chaitanya K. Joshi, Sékou-Oumar Kaba, Félix Therrien, Alex Hernandez-Garcia, Rocío Mercado, N. M. Anoop Krishnan, and Alexandre Duval. LeMat-GenBench: A unified evaluation framework for crystal gen- erative models.arXiv preprint...

work page arXiv 2025
[20]

Inverse design in search of materials with target functionalities.Nature Reviews Chem- istry, 2(4):0121, 2018

Alex Zunger. Inverse design in search of materials with target functionalities.Nature Reviews Chem- istry, 2(4):0121, 2018

2018
[21]

Has generative artificial intelligence solved inverse ma- terials design?Matter, 7(7):2355–2367, 2024

Hyunsoo Park, Zhenzhu Li, and Aron Walsh. Has generative artificial intelligence solved inverse ma- terials design?Matter, 7(7):2355–2367, 2024

2024
[22]

Bartel, and Gerbrand Ceder

Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Gerbrand Ceder. Chgnet as a pretrained univer- sal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence, 5(9):1031–1041, 2023

2023
[23]

Grossman

Tian Xie and Jeffrey C. Grossman. Crystal graphconvolutionalneuralnetworksforanaccurate and interpretable prediction of material properties. Physical Review Letters, 120(14):145301, 2018

2018
[24]

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

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

2019
[25]

Integrating electronic structure into generative modeling of inorganic materials.arXiv preprint arXiv:2511.14228, 2025

Junkil Park, Junyoung Choi, and Yousung Jung. Integrating electronic structure into generative modeling of inorganic materials.arXiv preprint arXiv:2511.14228, 2025

work page arXiv 2025
[26]

Richard F. W. Bader.Atoms in Molecules: A Quantum Theory. Oxford University Press, Oxford, 1990

1990
[27]

A fast and robust algorithm for Bader de- composition of charge density.Computational Ma- terials Science, 36(3):354–360, 2006

GraemeHenkelman, AndriArnaldsson, andHannes Jónsson. A fast and robust algorithm for Bader de- composition of charge density.Computational Ma- terials Science, 36(3):354–360, 2006

2006
[28]

Kenny, Roger Smith, and Graeme Henkelman

Edward Sanville, Steven D. Kenny, Roger Smith, and Graeme Henkelman. Improved grid-based algo- rithm for Bader charge allocation.Journal of Com- putational Chemistry, 28(5):899–908, 2007

2007
[29]

W. Tang, E. Sanville, and G. Henkelman. A grid-based Bader analysis algorithm without lat- tice bias.Journal of Physics: Condensed Matter, 21(8):084204, 2009

2009
[30]

Min Yu and Dallas R. Trinkle. Accurate and effi- cient algorithm for Bader charge integration.The Journal of Chemical Physics, 134(6):064111, 2011

2011
[31]

Mas- tej, and Aron Walsh

Masahiro Negishi, Hyunsoo Park, Kinga O. Mas- tej, and Aron Walsh. Continuous uniqueness and novelty metrics for generative modeling of inorganic crystals. InAI for Accelerated Materials Design - NeurIPS 2025, 2025

2025
[32]

A general-purpose machine learning framework for predicting properties of in- organic materials.npj Computational Materials, 2:16028, 2016

Logan Ward, Ankit Agrawal, Alok Choudhary, and Christopher Wolverton. A general-purpose machine learning framework for predicting properties of in- organic materials.npj Computational Materials, 2:16028, 2016

2016
[33]

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

Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table.Nature Computational Science, 2(11):718– 728, 2022

2022
[34]

Ilyes Batatia, Dávid Peter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. InAdvances in Neural Information Processing Sys- tems, volume 35, pages 11423–11436, 2022

2022
[35]

MatterSim: A Deep Learning Atomistic Model Across Elements, Temperatures and Pressures

Han Yang, Chenxi Hu, Yichi Zhou, Xixian Liu, Yu Shi, Jielan Li, Guanzhi Li, Zekun Chen, Shuizhou Chen, Claudio Zeni, Matthew Horton, Robert Pinsler, Andrew Fowler, Daniel Zügner, Tian Xie, Jake Smith, Lixin Sun, Qian Wang, Lingyu Kong, Chang Liu, Hongxia Hao, and Ziheng Lu. MatterSim: A deep learning atomistic model across elements, temperatures and press...

work page internal anchor Pith review arXiv 2024
[36]

Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations.Journal of Chemical Theory and Computation, 20(11):4857–4868, 2024

Yutack Park, Jaesun Kim, Seungwoo Hwang, and Seungwu Han. Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations.Journal of Chemical Theory and Computation, 20(11):4857–4868, 2024

2024
[37]

Mas- tej, and Aron Walsh

Masahiro Negishi, Hyunsoo Park, Kinga O. Mas- tej, and Aron Walsh. xtalmet: Python package containing distance functions and evaluation met- rics for crystal generation. GitHub repository, 2025. Accessed 2026-04-07

2025
[38]

Davies, Keith T

Daniel W. Davies, Keith T. Butler, Adam J. Jack- son, Jonathan M. Skelton, Kazuki Morita, and Aron Walsh. Smact: Semiconducting materials by analogy and chemical theory.Journal of Open Source Software, 4(38):1361, 2019. 12

2019
[39]

Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, ShreyasCholia, DanGunter, DavidSkinner, Gerbrand Ceder, and Kristin A. Persson. Commen- tary: The materials project: A materials genome approach to accelerating materials innovation.APL Materials, 1(1):011002, 2013

2013
[40]

Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg

Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. Struc- tured denoising diffusion models in discrete state- spaces. InAdvances in Neural Information Process- ing Systems, volume 34, 2021

2021
[41]

GemNet: Universal directional graph neural networks for molecules

Johannes Gasteiger, Florian Becker, and Stephan Günnemann. GemNet: Universal directional graph neural networks for molecules. InAdvances in Neural Information Processing Systems, volume 34, pages 6790–6802, 2021

2021
[42]

Generative model- ing by estimating gradients of the data distribution

Yang Song and Stefano Ermon. Generative model- ing by estimating gradients of the data distribution. InAdvances in Neural Information Processing Sys- tems, volume 32, 2019

2019
[43]

Chevrier, Kristin A

Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L. Chevrier, Kristin A. Persson, and Gerbrand Ceder. Python materials genomics (pymatgen): A robust, open- source python library for materials analysis.Com- putational Materials Science, 68:314–319, 2013

2013
[44]

Jonathan Schmidt, Noah Hoffmann, Hai-Chen Wang, Pedro Borlido, Pedro J. M. A. Carriço, Tiago F. T. Cerqueira, Silvana Botti, and Miguel A. L. Marques. Machine-learning-assisted deter- mination of the global zero-temperature phase diagram of materials.Advanced Materials, 35(22):2210788, 2023

2023
[45]

Efficient it- erative schemes forab initiototal-energy calcula- tions using a plane-wave basis set.Physical Review B, 54(16):11169–11186, 1996

Georg Kresse and Jürgen Furthmüller. Efficient it- erative schemes forab initiototal-energy calcula- tions using a plane-wave basis set.Physical Review B, 54(16):11169–11186, 1996

1996
[46]

Efficiency ofab-initiototal energy calculations for metals and semiconductors using a plane-wave basis set.Com- putational Materials Science, 6(1):15–50, 1996

Georg Kresse and Jürgen Furthmüller. Efficiency ofab-initiototal energy calculations for metals and semiconductors using a plane-wave basis set.Com- putational Materials Science, 6(1):15–50, 1996

1996
[47]

Perdew, Kieron Burke, and Matthias Ernz- erhof

John P. Perdew, Kieron Burke, and Matthias Ernz- erhof. Generalized gradient approximation made simple.Physical Review Letters, 77(18):3865–3868, 1996

1996
[48]

From ultrasoft pseudopotentials to the projector augmented-wave method.Physical Review B, 59(3):1758–1775, 1999

Georg Kresse and Daniel Joubert. From ultrasoft pseudopotentials to the projector augmented-wave method.Physical Review B, 59(3):1758–1775, 1999

1999
[49]

Mosca, Angeles Pulido, Andrew I

Daniel Widdowson, Marco M. Mosca, Angeles Pulido, Andrew I. Cooper, and Vitaliy Kurlin. Av- erage minimum distances of periodic point sets. MATCH Communications in Mathematical and in Computer Chemistry, 87(3):529–559, 2022. Data and code availability The code developed for this study will be made publicly available upon publication Acknowledgements This ...

2022