Source-linked AI summary
High-throughput discovery of novel cubic crystal materials using deep generative neural networks
Yong Zhao, Mohammed Al-Fahdi, Ming Hu, Edirisuriya MD Siriwardane, Yuqi Song, Alireza Nasiri, Jianjun Hu
TL;DR
Materials screening is limited by the quantity and diversity of structures in existing repositories, while current generative models struggle to produce diverse, chemically valid, and stable materials. CubicGAN uses a GAN trained on OQMD cubic materials to generate structures conditioned on elements and space groups. It rediscovered known cubic materials and produced new prototypes, including 506 materials verified as stable by DFT-based phonon dispersion checks.
Problem
Existing materials repositories have limited quantity and diversity, and current generative models struggle to generate structurally diverse, chemically valid, and stable materials.
Method
CubicGAN is a GAN trained on OQMD materials that generates cubic crystal structures conditioned on selected elements and space groups.
Results
506 new-prototype materials were generated and confirmed stable by phonon dispersion calculations.
Takeaways & Limitations
CubicGAN provides a route to expanding materials repositories with cubic structures of new composition prototypes for subsequent screening.
Takeaways & Limitations
The study focuses on ternary and quaternary cubic structures from three space groups: 216, 221, and 225.
Abstract
from arXiv · showhide
High-throughput screening has become one of the major strategies for the discovery of novel functional materials. However, its effectiveness is severely limited by the lack of quantity and diversity of known materials deposited in the current materials repositories such as ICSD and OQMD. Recent progress in machine learning and especially deep learning have enabled a generative strategy that learns implicit chemical rules for creating chemically valid hypothetical materials with new compositions and structures. However, current materials generative models have difficulty in generating structurally diverse, chemically valid, and stable materials. Here we propose CubicGAN, a generative adversarial network (GAN) based deep neural network model for large scale generation of novel cubic crystal structures. When trained on 375,749 ternary crystal materials from the OQMD database, we show that our model is able to not only rediscover most of the currently known cubic materials but also generate hypothetical materials of new structure prototypes. A total of 506 such new materials (all of them are either ternary or quarternary) have been verified by DFT based phonon dispersion stability check, several of which have been found to potentially have exceptional functional properties. Considering the importance of cubic materials in wide applications such as solar cells and lithium batteries, our GAN model provides a promising approach to significantly expand the current repository of materials, enabling the discovery of new functional materials via screening. The new crystal structures finally verified by DFT are freely accessible at our Carolina Materials Database http://www.carolinamatdb.org.
1 Introduction
Existing-material screening is constrained by limited repository scale and diversity, motivating generative approaches that create broader crystal-structure candidates. CubicGAN conditions a GAN on elements and space groups, recovering known cubic materials while producing new prototypes subsequently validated through DFT.
- Motivation: Existing repositories contain far fewer structures than the nearly infinite chemical design space, limiting screening-based materials discovery.ICSD and MP contain about 165,000 and 125,000 materials, respectively.
- Related approaches: Element substitution remains constrained by the limited number of known prototypes, while generative models offer a route to broader composition and structure exploration.Prior approaches include element substitution, composition generation with structure prediction, and generative machine learning models.
- Approach: CubicGAN generates cubic structures by conditioning a GAN on selected elements and a specified space group, using lattice parameters, atom coordinates, element embeddings, and space-group information.The training set consists of ternary materials selected from OQMD because of its size and compositional diversity.
- Results: CubicGAN rediscovered many known cubic structures and identified new composition prototypes, including ABC6-216, ABC6D6-216, and AB8C12-221.The experiments compare generated materials against existing databases and assess their stability.
- Approach: 10 million ternary and 10 million quaternary hypothetical crystal structures were generated for downstream analysis.The model targets large-scale generation of cubic materials for subsequent screening and validation.
- Results: 506 new-prototype materials were generated and confirmed stable by phonon dispersion calculations.A preceding DFT analysis examined 108,897 hypothetical materials, with 33.8% of novel materials successfully relaxed.
2 Methods
CubicGAN is trained on selected cubic ternary and quaternary OQMD structures, using space-group and elemental conditions to generate materials and validation datasets to assess rediscovery. Its framework combines conditioned GAN generation with structural discrimination, Wasserstein training, and lattice-parameter post-processing.
- Dataset: The study selects ternary and quaternary cubic structures from OQMD, focusing on three space groups that cover most known cubic materials.The selected space groups are 216, 221, and 225; quaternary training is restricted to space group 216 because of strong dataset imbalance.
- Dataset: The training representation retains three ternary or four quaternary nonequivalent atom positions, enabling unified matrix representations and symmetry-based reconstruction of full structures.For ternary materials, the representation uses a (28 × 3) matrix; quaternary materials use a (27 × 3) matrix in the selected setting.
- Dataset: 375,749 ternary materials from Fm¯3m, F¯43m, and Pm¯3m form the OQMD-TC3 training dataset, representing 249,646 unique formulas.The dataset uses 84 elements and is designed to learn valid ternary element combinations.
- Dataset: Materials Project and ICSD cubic ternary structures provide held-out validation datasets for measuring rediscovery rates.The validation collection contains 6,545 materials from MP and 1,875 from ICSD, excluded from training.
- CubicGAN Framework: The generator receives a space group, three elements, and random noise, while the discriminator evaluates coordinates, element properties, unit-cell parameters, and space group.Element and space-group inputs are embedded before generation, and symmetry determines the number of atoms.
- CubicGAN Framework: Wasserstein distance with gradient penalty stabilizes GAN training, while a composition-based model post-processes the generated cubic lattice parameter.The gradient-penalty balancing parameter λ is set to 10, and the lattice-parameter predictor reports R2 = 0.979 for cubic lattice a prediction.
3 Results and Discussion
CubicGAN efficiently rediscovers known cubic materials and generates chemically novel prototypes, which are filtered through staged stability checks. DFT phonon calculations confirm 506 stable new-prototype materials, while structural embeddings show distinct distributions from known materials.
- Rediscovery rate: 10 million samples rediscovered 95.5% of the OQMD-TC3 training materials, 72.0% of MP-TC3 materials, and 50.7% of ICSD-TC3 materials.The lower validation-set rates reflect different space-group proportions; half of the rediscovered MP-TC3 materials were stable by formation-energy and energy-above-hull criteria.
- Sampling efficiency: 44,507 was the enrichment score for generating chemically valid ternary structures compared with exhaustive enumeration.The score used 95.5% rediscovery of OQMD-TC3 from 10 million samplings against 466,055,331,840 possible configurations.
- Stability validation: Three validation stages reduced generated candidates before DFT calculations, combining chemical filtering, structural relaxation, and mechanical and vibrational stability checks.Phonon dispersion was the eventual stability criterion after formation-energy prediction, DFT optimization, Γ-point frequencies, and elastic-constant filtering.
- Novel prototype generation: 24 ternary and 1 quaternary novel prototypes were identified among filtered generated materials, including 209,744 ternary and 260,891 quaternary new-prototype candidates.Lanthanide- and actinide-containing samples were removed before this prototype analysis.
- DFT-confirmed stability: 506 new-prototype materials were confirmed stable by phonon dispersion: 183 ternary and 323 quaternary materials across four prototypes.The stable prototypes were ABC6-216, AB6C6-225, ABCD6-216, and ABC6D6-216; ABC6-216 and ABC6D6-216 were novel relative to the training and validation sets.
- Potential functional properties: Phonon dispersions revealed tunable bandgaps, soft acoustic modes, and steep optical-mode gradients with possible relevance to photovoltaics, superconductors, and thermoelectrics.These observations were presented as potential functional opportunities requiring further investigation.
- Structural diversity: t-SNE embeddings of 901-dimensional XRD representations showed new-prototype materials forming structurally distinct clusters from known cubic materials.For some prototypes, new materials were peripheral to known clusters; others formed multiple distinct clusters containing DFT-verified stable materials.
4 Conclusion
CubicGAN generates cubic crystal structures at large scale, rediscovering known structures while producing structurally distinct new prototypes. DFT relaxation and phonon-dispersion calculations verified 506 new-prototype cubic materials.
- 375,749 ternary cubic crystal structures from OQMD trained CubicGAN to generate three major cubic space groups: 216, 225, and 221.
- Within 10 million samplings, CubicGAN rediscovered most known cubic structures and generated materials with new structure prototypes.
- 24 new cubic-material prototypes were identified through further analysis of the generated structures.
- 506 new-prototype cubic materials were verified using DFT-based relaxation and phonon-dispersion calculations.
5 Contribution
The paper credits its authors across conceptualization, methodology, software, validation, investigation, writing, visualization, supervision, and funding.
- Authors contributed across the study’s conceptualization, methodology, software, validation, investigation, and data curation.The listed contributors also covered resources, writing, visualization, supervision, and funding acquisition.