Symmetry-aware generative design of flat-band materials beyond known crystal-net prototypes

This paper introduces SkeleGen, a symmetry-constrained generative model that expands the discovery of flat-band materials beyond traditional geometric prototypes by generating and validating over 9,000 stable crystal candidates featuring novel, out-of-distribution sublattice motifs.

The Gist

Imagine you are trying to build a new type of super-efficient electronic device. To do this, you need a special kind of material where electrons get "stuck" in place, moving very slowly. In physics, we call these "flat bands." When electrons are stuck, they interact strongly with each other, creating cool quantum effects like superconductivity.

For a long time, scientists have been trying to find these materials by looking at a very small, fixed list of geometric shapes. It's like trying to build a house using only three specific blueprints: a Kite shape, a Star shape, and a Pyramid shape. If your house doesn't fit one of those three blueprints, scientists assumed it couldn't have the "stuck electron" property.

The Problem: The "Catalog" is Too Small
The authors of this paper realized that this list of three shapes is way too short. Real crystals are messy and complex. There are thousands of other geometric patterns that could trap electrons, but because they don't look exactly like the famous "Kite" or "Star" shapes, they were ignored. They were invisible to the standard design tools.

The Solution: A "Skeleton" Generator
The team developed a new AI tool called SkeleGen. Think of it like a master architect who doesn't just copy old blueprints but can invent entirely new ones.

Here is how their process works, step-by-step:

1. Finding the "Skeleton": First, they looked through a massive database of existing materials (like a giant library of crystals). They stripped away all the extra atoms and chemicals, leaving behind just the bare "skeleton" of the structure—the framework of connections between atoms. They found 1,852 of these skeletons.
2. Spotting the "Outsiders": Most of these skeletons looked like the famous shapes everyone knows. But they found 236 "outliers." These were weird, unique skeletons that didn't fit any known category. They were the "unknown shapes" that the old catalog missed.
3. The "Invisible Scaffold": The team treated these weird skeletons as rigid, unchangeable scaffolds. Imagine you have a strange, twisted wire frame (the skeleton). You can't bend it or break it; it must stay exactly as it is.
4. Filling in the Blanks: This is where the AI shines. The AI's job is to figure out what chemicals (atoms) to put around that rigid wire frame to make a stable, real-world crystal. It's like asking, "If I have this weird wire frame, what kind of bricks, mortar, and paint can I use to build a house that stands up?"
5. Symmetry Check: The AI is very careful to ensure that the new chemicals it adds don't break the symmetry of the wire frame. If the frame is symmetrical, the house must be too, or the "stuck electron" effect will disappear.

The Results
The AI generated over one million potential new materials based on these weird skeletons. After filtering out the ones that wouldn't be stable, they were left with 9,352 promising candidates.

They then used powerful supercomputers to simulate the physics of these materials. The results were amazing:

• 73% of the tested materials actually had the "flat band" property they were looking for.
• Crucially, the "stuck electron" effect came directly from the weird wire frame they started with. It wasn't a lucky accident; the geometry caused the effect.

The Big Picture
The paper claims that by using these "out-of-distribution" skeletons (shapes that don't fit the old rules), they have unlocked a massive new design space. They proved that you don't need to stick to the famous "Kite" or "Star" shapes to find these special materials. You can use any geometric pattern that creates the right connectivity, and the AI can figure out how to build a real crystal around it.

In short: They stopped looking for materials that fit old boxes and started building new boxes around materials that were previously ignored.

Technical Summary

Technical Summary: Symmetry-aware generative design of flat-band materials beyond known crystal-net prototypes

Problem Statement

The design of materials hosting flat electronic bands—crucial for strongly correlated phenomena like unconventional superconductivity and fractional Chern insulators—has historically relied on a discrete catalogue of named geometric motifs (e.g., kagome, Lieb, pyrochlore). This approach is limited because:

1. Discrete Taxonomy: Standard crystal net taxonomies partition geometries into discrete prototypes, failing to identify or classify structures that fall between these named families.
2. Invisibility of Novelty: Materials with flat bands arising from connectivity patterns outside these named families are effectively invisible to current design strategies, even if they exist in structural databases.
3. Limitations of Generative Models: While recent diffusion-based generative models (e.g., CDVAE, DiffCSP, MatterGen) have improved structural validity and property control, they generally do not enforce the retention of specific, uncatalogued local connectivity patterns. Existing symmetry-aware methods either constrain global degrees of freedom or treat geometric templates as structural constraints rather than preserving their space-group-consistent Wyckoff configurations.

Consequently, there is a lack of a framework to transfer "uncatalogued" flat-band connectivity patterns into new, chemically realizable crystals while preserving both their symmetry and electronic origin.

Methodology: The SkeleGen Framework

The authors propose SkeleGen, a skeleton-guided generative framework that treats unconventional flat-band geometries as transferable design variables. The workflow consists of four main stages:

1. Identification of Connectivity-Driven Flat-Band Skeletons

• Screening: Starting from the Materials Project (MP) database, 29,465 materials were screened using a physics-motivated flatness score (combining bandwidth and density of states). 4,155 materials with high flatness scores were retained.
• Sublattice Abstraction: For high-scoring materials, elemental sublattices were extracted and converted into nearest-neighbor crystal graphs. A single $s$ orbital was assigned to each site to construct a tight-binding (TB) Hamiltonian.
• Filtering: Sublattices were retained only if their nearest-neighbor Hamiltonian supported a dispersionless band after removing isolated finite clusters. This yielded 1,852 non-trivial flat-band skeletons where flatness is confirmed to originate from network connectivity rather than trivial atomic insulation.

2. Geometric Latent Space Analysis and Novelty Scoring

• Continuous Representation: To move beyond categorical classification, a Graph Neural Network (GNN) encoder was trained via contrastive self-supervised learning on the 1,852 skeletons. The encoder uses only Cartesian coordinates as node features, producing geometric embeddings independent of chemistry.
• Clustering and Classification:
  • Known Prototypes: A barycentric placement algorithm (Systre) assigned RCSR names to only 334 skeletons, highlighting the limitations of named-net classification for distorted or low-symmetry structures.
  • Recurrent Geometries: HDBSCAN clustering grouped 1,295 skeletons into recurrent geometric clusters, many of which lacked catalogue names.
• Novelty Quantification: A novelty score ($S_{novelty}$) was defined as the minimum latent-space distance to any point classified by Systre or clustered by HDBSCAN.
• Selection: Skeletons with near-zero scores (within known regions) were discarded. 236 high-novelty skeletons (above the 50th percentile, $S_{novelty} > 0.27$) were selected. These represent geometries that are neither named nor clustered, spanning diverse graph-theoretic descriptors (e.g., coordination numbers, odd-cycle fractions) distinct from known families.

3. Symmetry-Aware Constrained Generation

The 236 novel skeletons were used as rigid scaffolds to generate new crystals:

• Wyckoff Constraints: The fixed fractional coordinates of the skeleton were mapped to symmetry-compatible Wyckoff positions within a sampled space group ($G_{full}$), which is a subgroup of the skeleton's original space group.
• Diffusion Process: The framework combines SCIGEN (structural constraint masking) with DiffCSP++ (symmetry-aware diffusion).
  • Constrained Atoms ($M^c_t$): The skeleton atoms follow a predefined forward-diffusion trajectory, ensuring the geometric core remains intact.
  • Unconstrained Atoms ($M^u_t$): The remaining atoms are denoised by the model $\phi_\theta$ to fill the chemical environment.
  • Recombination: At each timestep, the full structure is recombined via a constraint mask.
• Adaptive Unmasking: In the final 10% of diffusion steps, constrained atoms transition from rigid masking to standard Wyckoff-constrained denoising, allowing for minor positional relaxation within the symmetry subspace to improve stability.

4. Screening and Validation

• Initial Filtering: 1,030,975 candidates were generated. A four-stage stability screen (SMACT charge balance, occupation ratio, energy-above-hull, non-diffused structure checks) reduced this to 72,544.
• Surrogate Filtering: A GNN surrogate model trained on flatness scores (bandwidth and DOS) pre-filtered the pool, reducing it to 9,352 candidates.
• High-Throughput DFT: Full Density Functional Theory (DFT) calculations (VASP, PBEsol) were performed on a random sample of 300 candidates to validate the pipeline.

Key Results

1. Expansion of the Geometric Repertoire

The analysis of the 1,852 skeletons revealed that named motifs (kagome, Lieb, etc.) describe only a small fraction of flat-band geometries in real crystals. The 236 selected high-novelty skeletons represent a diverse set of "out-of-distribution" geometries, including 3D expansions of 2D nets, motifs obscured by complex local environments, and entirely new connectivity patterns (e.g., the Fe sublattice in $Fe_7S_8$).

2. Successful Generation of Flat-Band Materials

• Success Rate: Among 300 randomly sampled candidates from the 9,352 screened pool, 256 converged. Of the 184 suitable for evaluation, 134 (73%) achieved a flatness score $> 0.5$, and 78 achieved $> 0.95$.
• Mechanistic Link: The flat bands in the generated crystals were confirmed to originate from the imposed skeletons. In representative cases (e.g., $KNaTlCoSbF_{12}$, $La(FeO_4)_2$, $TaAsSF_{15}$), the DFT band structures closely matched the tight-binding spectra of the isolated skeletons, and the flat spectral weight was concentrated on the constrained sublattice atoms.
• Symmetry Preservation: The generated materials retained the crystallographic symmetry required for the flat-band features, with band touching points occurring at high-symmetry points, suggesting potential for topologically non-trivial flat bands.

3. Representative Candidates

The study identified specific new materials, such as:

• $KNaTlCoSbF_{12}$: Features a 2D F-based net extended into a 3D crystal, with a flat band ~1.4 eV below the Fermi level.
• $La(FeO_4)_2$: Built around an O-based corner-sharing tetrahedral framework, exhibiting a manifold of flat bands 0.8 eV below the Fermi level.
• $TaAsSF_{15}$: Contains a complex 3D F-based skeleton with no known counterpart, recovering flat features directly at the Fermi level.

Significance and Claims

The paper claims to establish a new design principle for materials discovery: "out-of-distribution" motifs.

• Beyond Search: The work demonstrates that flat-band materials can be discovered not by searching existing databases for known motifs, but by generating new crystals from uncatalogued connectivity patterns.
• Mechanism-Preserving Transfer: The framework successfully transfers the geometric mechanism of flat bands (destructive interference via connectivity) into new chemical environments while respecting crystallographic symmetry. This addresses the "model-to-material" gap by embedding abstract lattice logic directly into crystal generation.
• Scalability: The approach scales to over one million candidates, proving that unconventional geometries can be embedded into a vast chemical search space.
• Future Potential: While the current implementation focuses on nearest-neighbor connectivity, the authors suggest the strategy could be extended to include multi-orbital character, spin-orbit coupling, and targeted topological invariants, moving from discovering flat-band candidates to designing specific correlated quantum phases.

The authors emphasize that their results show the geometric origins of flat bands are not exhausted by the established catalogue of named motifs, revealing a much broader landscape of connectivity-driven flat-band materials.