🔬 materials science

Diffusion-Based Generative Screening of ABX $_3$ Perovskite Compositions for Bandgap-Targeted Materials Discovery

This study evaluates a pretrained diffusion model for generating ABX $_3$ perovskite candidates with targeted bandgaps, demonstrating its ability to explore diverse chemical compositions and ordered structures while highlighting critical limitations in predictive reliability and the necessity of rigorous structural and property validation before treating generated outputs as viable materials.

Original authors: Sunraj Patel, Krishanu Biswas

Published 2026-11-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sunraj Patel, Krishanu Biswas

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Materials scientists have long been fascinated by a family of crystals known as perovskites. These materials are not just a single substance but a vast architectural family where atoms arrange themselves into a specific, repeating three-dimensional pattern. Imagine a framework built from tiny octahedrons—shapes with eight faces—that share their corners to form a rigid, interconnected skeleton. Within the spaces of this skeleton sit other atoms that can be swapped out like pieces in a complex puzzle. This flexibility allows researchers to tune the material's properties, such as how it conducts electricity or interacts with light, making perovskites prime candidates for next-generation solar cells and electronic devices. However, the sheer number of possible combinations of atoms is so enormous that trying to find the perfect one by hand is like searching for a needle in a haystack the size of a city.

To tackle this challenge, researchers are increasingly turning to artificial intelligence to generate new crystal structures on the computer. These digital models can propose millions of atomic arrangements in a fraction of the time it would take a human to sketch them out. The critical question, however, is whether these computer-generated ideas are actually real, stable materials or just mathematical ghosts that look good on paper but fall apart when examined closely. A new study by Sunraj Patel and Krishanu Biswas at the Indian Institute of Technology Kanpur puts this technology to the test specifically for perovskites. They wanted to see if a powerful AI tool could not only invent new versions of these crystals but also reliably predict which ones would have the specific electronic properties needed for practical use, such as a particular energy gap that allows a material to absorb sunlight efficiently.

The researchers used a sophisticated AI system called a diffusion model, which works by learning the patterns of known crystals and then gradually shaping random noise into new, plausible structures. They instructed this system to look for crystals with a specific electronic property: a bandgap of 1.4 electronvolts, a value often associated with high-performance solar energy conversion. The AI generated thousands of candidate structures, but the team knew that simply matching the chemical formula was not enough. In the world of crystals, the same recipe of atoms can sometimes assemble into completely different shapes, some of which lack the essential corner-sharing framework that defines a true perovskite. To filter out these imposters, the team applied a strict connectivity test, checking if the atoms actually formed the required three-dimensional network of linked octahedrons.

This structural filter proved to be a crucial gatekeeper. Out of 3,500 crystals generated by the AI that had the correct chemical ratio of atoms, only about 198 actually possessed the correct three-dimensional framework. After removing duplicates and checking for chemical stability, the team was left with 171 unique, valid perovskite candidates. This result highlighted a significant limitation of looking only at chemical formulas: without checking the actual atomic connections, a search would have accepted more than twice as many structures that were not true perovskites at all. The study also revealed that the AI was particularly good at inventing complex variations where the central part of the crystal structure was occupied by more than one type of atom, a feature known as mixed-site occupancy. These complex arrangements, which are difficult to create using traditional methods of swapping atoms one by one, were a major strength of the generative approach.

However, the study also exposed the boundaries of what current AI can reliably predict. While the AI successfully generated a diverse set of structures, its ability to accurately forecast their electronic properties was mixed. When the researchers tested the AI's predictions against a large database of known materials, the model performed well for familiar chemical combinations. But when they challenged it with a test involving rare or unfamiliar atoms that the model had never seen during its training, its predictions became no better than a random guess. This suggests that while the AI is excellent at recombining known elements into new patterns, it struggles to extrapolate to entirely new chemical territories. Furthermore, when the team took a small selection of the AI's best candidates and performed rigorous physics-based calculations to see how the atoms would settle into their lowest energy state, the results were surprising. In some cases, the atoms shifted slightly during this relaxation process, changing the crystal's classification or altering its predicted electronic properties significantly.

The most striking finding was that none of the six generated structures the team tested with high-precision calculations ended up having the exact target property the AI had been asked to find. In several instances, the AI's initial prediction of the material's energy gap was off by a noticeable margin, and the gap changed again after the atoms were allowed to relax into their natural positions. This does not mean the AI failed completely; it successfully identified a rich landscape of new structural possibilities that traditional methods might have missed. But it does show that the path from a computer-generated idea to a confirmed, useful material is still fraught with uncertainty. The study concludes that while these generative models are powerful tools for exploring the vast chemical space of perovskites, they cannot yet replace the need for careful, step-by-step verification. The AI can propose the candidates, but human scientists must still act as the final judges, checking the structural integrity and calculating the true properties before declaring a discovery.

Technical Summary: Diffusion-Based Generative Screening of ABX3 Perovskite Compositions

Problem Statement
The discovery of functional perovskite materials (ABX3) is hindered by a vast search space where composition does not uniquely determine topology. Conventional enumeration strategies often rely on fixed prototypes (e.g., single-B site substitutions), which fail to capture the diversity of mixed-B and ordered double-perovskite configurations (A2BB'X6). Furthermore, generative models can produce structures with correct stoichiometry (3:2 anion-to-cation ratio) that lack the defining three-dimensional corner-sharing octahedral network of perovskites. Additionally, while machine learning (ML) predictors can estimate properties like bandgaps, their reliability often degrades when extrapolating to unfamiliar chemical spaces (e.g., disjoint B-site elements), and first-principles relaxation can alter structural classifications and property values significantly.

Methodology
The study employs a multi-stage screening pipeline to evaluate the utility of a pretrained diffusion model (MatterGen) for targeted materials discovery:

Crystal Generation: Periodic crystals were generated using the MatterGen model conditioned on a target bandgap of 1.4 eV (using classifier-free guidance strengths $\gamma=2$ and $\gamma=4$ ) and an unconditional checkpoint. A reference set of 4,000 structures was also constructed via substitution into cubic SrTiO3 and orthorhombic CaTiO3 prototypes to serve as a baseline for accessible structural classes.
Structural Filtering: Generated structures underwent rigorous filtering:
- Compositional: Ordered structures with a 3:2 anion-to-cation ratio.
- Topological: A connectivity-based classifier identified 3D corner-sharing BX6 octahedral networks. Cations were assigned as B-sites if they had six anion neighbors within a specific distance threshold, and the resulting network was verified for 3D periodicity.
- Chemical: Charge balance and Goldschmidt tolerance factors (0.78–1.10) were applied where data permitted. Duplicates were removed using pymatgen's StructureMatcher.
Property Prediction: An ensemble of ALIGNN (Atomistic Line-Graph Neural Network) models, trained on 32,058 curated Materials Project (MP) structures, predicted bandgaps and energies above the convex hull.
Validation:
- Predictor Performance: Evaluated on two test splits: a formula-grouped random split and a "B-site-disjoint" split (excluding 14 specific elements from training) to test chemical transferability.
- First-Principles Checks: Six generated structures and two controls (SrTiO3, KNbO3) were relaxed using Density Functional Theory (DFT) with Quantum ESPRESSO (PBE functional). Structural retention and bandgap comparisons between ALIGNN predictions and DFT calculations were analyzed.

Key Contributions

Topology-Aware Screening: Demonstrates that stoichiometry alone is insufficient for identifying perovskites; a connectivity-based filter is essential to distinguish true 3D corner-sharing frameworks from non-perovskite arrangements sharing the same nominal formula.
B-Site Multiplicity: Reveals that diffusion-based generation accesses mixed-B and ordered double-perovskite configurations (69% of accepted structures) that are inaccessible to single-prototype substitution methods.
Predictor Limitations: Quantifies the performance gap of ML property predictors. While ALIGNN performs well on random splits (MAE 0.25 eV), it fails to outperform a constant baseline on B-site-disjoint tests (MAE 0.55 eV vs. 0.53 eV for constant zero), highlighting limited transferability to unseen chemistries.
Relaxation Effects: Shows that structural relaxation can alter the classification of a generated structure (e.g., TlPbCl3 transforming from perovskite to non-perovskite due to A-site coordination loss) and significantly shift bandgap values, sometimes reversing the sign of the prediction error relative to DFT.

Results

Yield and Diversity: From 3,500 conditioned samples ( $\gamma=2$ ), only 198 (5.7%) satisfied the topological perovskite definition, yielding 171 unique structures. The generated set showed a higher fraction of halide-containing compositions (45%) compared to the MP reference (22%) and a higher fraction of mixed-B structures (69%) compared to the substitution reference (0%).
Target Window: The fraction of generated perovskites falling within the target window (1.4 $\pm$ 0.3 eV) was 8.2%, comparable to the MP reference fraction (7.3%). However, the median predicted gap shifted closer to the target (1.48 eV) compared to the unconditional set (0.78 eV).
Database Novelty: Of the 171 unique structures, 134 matched existing entries in MP or Alex-MP-20. The unmatched fraction (37 structures) did not yield any candidates that met all screening criteria (gap, hull, and ensemble spread) while being unmatched from both databases.
DFT Verification: Among the six generated structures relaxed via DFT, five retained the perovskite framework. Three had PBE gaps within the target window (1.1–1.7 eV), but none were simultaneously within the window according to both ALIGNN (on starting geometry) and PBE. ALIGNN tended to overpredict gaps for halides relative to PBE on relaxed geometries.

Significance and Claims
The paper concludes that crystal diffusion models offer a viable route for exploring mixed-site and ordered perovskite configurations beyond narrow prototype sets, effectively recombining known elements into new site arrangements. However, the study modestly claims that these results do not establish the discovery of new stable compounds or successful inverse design for the specific 1.4 eV target.

The authors emphasize that:

Connectivity is critical: Structural family recovery must be assessed via connectivity, not just stoichiometry.
Predictor bias exists: Downstream property models (like ALIGNN) require careful validation, particularly for disjoint chemical spaces, and their outputs should not be treated as definitive without first-principles checks.
Relaxation matters: Initial geometry predictions can differ substantially from relaxed-state properties, and relaxation can fundamentally change the structural classification.

The study positions itself as a necessary "structural and predictive check" before generative outputs can be treated as viable materials candidates, rather than a demonstration of superior general validity or stability compared to existing benchmarks.

Technical Summary: Diffusion-Based Generative Screening of ABX3 Perovskite Compositions

More like this