Diffusion-Based Generative Screening of ABX3 Perovskite Compositions for Bandgap-Targeted Materials Discovery
Abstract
Crystal-generation models can explore compositions and atomic arrangements beyond a fixed set of crystal prototypes, but their usefulness for a specific materials family depends on whether the generated structures retain the defining topology and whether their predicted properties remain reliable across the sampled chemistry. Here, a pretrained diffusion model is used to generate candidate ABX3 perovskites and to examine their structural, chemical and electronic-property distributions. A connectivity-based filter distinguishes three-dimensional corner-sharing perovskites from structures that share the same nominal ABX3 stoichiometry. The accepted set contains substantial B-site multiplicity, including ordered double-perovskite configurations that are inaccessible to the two-prototype substitution reference used here. The generated set is enriched in chloride-, bromide- and iodide-containing compositions relative to the Materials Project, while all generated B-site elements are already represented in that database. On the downstream MP-trained ALIGNN scale, the predicted bandgap distribution is shifted towards 1.4 eV, but the target-window fraction (8.2\%) is close to the MP reference fraction (7.3\%); the separate MatterGen and ALIGNN label sources prevent this comparison from establishing target-condition enrichment. An ALIGNN model performs well on a formula-grouped random test split but does not outperform a constant baseline on a deliberately severe B-site-disjoint test. First-principles calculations on six generated structures show that relaxation can alter structural classification and that predicted and calculated gaps can differ appreciably. These results identify both the chemical-space coverage offered by crystal diffusion and the structural and predictive checks needed before its outputs can be treated as materials candidates.
Related
- Statistical validation of calorimeter inpainting with generative diffusion priors
- Topology-Stratified Materials Discovery with A Flow-Based Generative Model
- A generative diffusion framework for physically consistent 3D turbulence
- Physics-Informed and Knowledge-Driven Generative AI for Autonomous Discovery of Porous Oxide Energy Materials: Opportunities and Challenges
- Some remarks on the connection between generative diffusion and optimal transport
BibTeX
@misc{patel2026diffusionbased,
title = {{Diffusion-Based Generative Screening of ABX$_3$ Perovskite Compositions for Bandgap-Targeted Materials Discovery}},
author = {Sunraj Patel and Krishanu Biswas},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.00766},
eprint = {2611.00766},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.00766},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Diffusion-Based Generative Screening of ABX_3 Perovskite Compositions for Bandgap-Targeted Materials Discovery
AU - Sunraj Patel
AU - Krishanu Biswas
PY - 2026
DA - 2026/11/01
PB - arXiv
DO - 10.48550/arxiv.2611.00766
UR - https://arxiv.org/abs/2611.00766
AB - Crystal-generation models can explore compositions and atomic arrangements beyond a fixed set of crystal prototypes, but their usefulness for a specific materials family depends on whether the generated structures retain the defining topology and whether their predicted properties remain reliable across the sampled chemistry. Here, a pretrained diffusion model is used to generate candidate ABX_3 perovskites and to examine their structural, chemical and electronic-property distributions. A connectivity-based filter distinguishes three-dimensional corner-sharing perovskites from structures that share the same nominal ABX_3 stoichiometry. The accepted set contains substantial B-site multiplicity, including ordered double-perovskite configurations that are inaccessible to the two-prototype substitution reference used here. The generated set is enriched in chloride-, bromide- and iodide-containing compositions relative to the Materials Project, while all generated B-site elements are already represented in that database. On the downstream MP-trained ALIGNN scale, the predicted bandgap distribution is shifted towards 1.4 eV, but the target-window fraction (8.2\%) is close to the MP reference fraction (7.3\%); the separate MatterGen and ALIGNN label sources prevent this comparison from establishing target-condition enrichment. An ALIGNN model performs well on a formula-grouped random test split but does not outperform a constant baseline on a deliberately severe B-site-disjoint test. First-principles calculations on six generated structures show that relaxation can alter structural classification and that predicted and calculated gaps can differ appreciably. These results identify both the chemical-space coverage offered by crystal diffusion and the structural and predictive checks needed before its outputs can be treated as materials candidates.
ER -