🔬 materials science

GLASS: Global Latent Aggregation with Slot-based Set Decoding for Scalable All-Atom Crystal Generation

The paper introduces GLASS, a generative model that overcomes the scalability challenges of all-atom crystal generation by decoupling correspondence assignment from generative transport through global latent aggregation and slot-based set decoding, enabling the high-validity creation of large-scale structures like metal-organic frameworks without relying on building block conditioning.

Original authors: Hendrik Kraß, Seyed Mohamad Moosavi, Mathias Niepert

Published 2026-09-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hendrik Kraß, Seyed Mohamad Moosavi, Mathias Niepert

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

Crystals are the ordered building blocks of the solid world, arranging their atoms into repeating patterns that determine everything from the hardness of a diamond to the ability of a battery to hold a charge. For decades, scientists have sought to design new crystals from scratch, hoping to discover materials that can capture carbon from the air, purify water, or store energy more efficiently. The challenge lies in the sheer vastness of the possibilities; the number of ways atoms can combine is so immense that useful structures are like needles in a cosmic haystack. To find them, researchers have turned to artificial intelligence, teaching computers to imagine and generate new atomic arrangements. However, as these digital experiments grow larger, aiming to create complex frameworks with hundreds of atoms rather than just a few, the computers often stumble. They struggle to keep the atoms in the right places, producing structures that collapse or simply do not exist in reality.

A team of researchers at the University of Stuttgart and the University of Toronto has identified the root of this stumbling block and proposed a new way to navigate it. They found that the difficulty arises not from a lack of computing power, but from a fundamental confusion in how the computer tracks individual atoms. When a model tries to generate a large crystal, it must decide which starting point moves to which final destination for every single atom. As the number of atoms increases, the number of possible ways to pair them up explodes, creating a chaotic landscape where the computer cannot learn a reliable path. The researchers call this the correspondence problem. To solve it, they developed a new system called GLASS, which stops trying to track every atom individually during the generation process. Instead, it first compresses the entire crystal into a single, unified summary, generates a new summary from scratch, and then unpacks that summary into a full atomic structure all at once.

The researchers tested this approach on two very different sets of data. First, they looked at small crystals containing up to twenty atoms, a standard size for previous models. Here, their new system performed just as well as the best existing methods, producing valid structures that matched the quality of known materials. But the true test came when they applied the system to metal-organic frameworks, a class of complex materials used for gas storage and filtration that can contain up to one hundred and fifty atoms in a single repeating unit. Previous attempts to generate these large structures without giving the computer a pre-defined blueprint often resulted in failures, with the generated structures falling apart or violating basic chemical rules. GLASS, however, succeeded where others struggled. It generated these large, complex frameworks with a success rate that approached the quality of the training data itself, producing valid structures even at the largest sizes tested.

The key to this success was a change in strategy. Rather than forcing the computer to learn a specific, fixed order for every atom—a task that becomes impossible as the system grows—the new method uses a set of learned "slots." Imagine these slots as empty seats at a table that the computer fills in simultaneously. The computer first creates a global picture of the crystal, a summary that captures the overall shape and composition without worrying about which specific atom goes where. It then uses this summary to fill the slots, deciding in parallel what type of atom belongs in each seat and where it should sit. This removes the need to match individual starting points to individual ending points, a step that previously caused the models to fail as the crystals got bigger. By separating the generation of the overall shape from the assignment of specific atoms, the system avoids the confusion that plagued earlier attempts.

Despite this technical triumph, the researchers found a significant limitation in the nature of what the computer was creating. While the system could reliably build large, valid structures, most of the new crystals it produced were not truly new. Instead, they were very close copies of the structures the computer had seen during its training. The system was excellent at reconstructing familiar frameworks but struggled to invent novel ones that had never been seen before. This suggests that while the new method solves the problem of building large structures, the ability to discover entirely new chemical spaces still depends on the diversity of the data the computer is fed. The researchers noted that the system's success in generating valid structures was high, but its ability to produce unique, never-before-seen materials remained low, indicating that the bottleneck has shifted from simply building the structure to imagining something truly different.

The implications of this work are clear for the future of materials science. The study demonstrates that the difficulty in generating large crystals is not an inherent flaw in the physics of the materials, but a specific hurdle in how the computer learns to organize them. By removing the need to track individual atom correspondences, the researchers have shown a path toward generating much larger and more complex systems than was previously possible. However, the findings also serve as a reminder that high validity does not automatically mean high novelty. The system can build the house, but it is still learning how to design a new kind of home. As the field moves forward, the focus will likely shift to improving the computer's ability to generalize from its training data, ensuring that the valid structures it builds are also the novel ones that scientists need to solve real-world problems.

Technical Summary: GLASS – Global Latent Aggregation with Slot-Based Set Decoding

1. Problem Statement

The paper addresses the challenge of scaling generative models for all-atom crystal structures to larger systems, such as Metal–Organic Frameworks (MOFs). While existing models can generate valid small crystals (e.g., up to 20 atoms), they struggle with larger systems, often exhibiting lower validity and requiring conditioning on building blocks, topology, or composition.

The authors identify the root cause as the correspondence problem in particle-space generation. In permutation-equivariant flow matching, a training path connects randomly placed source particles to target atoms. Since atomistic sets lack intrinsic ordering, defining a path requires pairing source particles to target atoms.

Ambiguity: Different pairings describe the same physical structure but prescribe different particle-wise trajectories.
Discontinuity: Optimal transport (OT) pairings switch discontinuously as source configurations change.
Equivariance Constraints: Deterministic equivariant models must produce identical outputs for particles in identical states. As set density increases, distinguishing nearby particles requires extreme sensitivity, making the learning of these discontinuous boundaries difficult.
Scaling Bottleneck: Experiments show that the training required for reliable generation in index-free permutation-equivariant flows increases substantially with set cardinality and density, even on a single fixed target.

2. Methodology: GLASS

To resolve the correspondence problem, the authors introduce GLASS (Global Latent Aggregation with Slot-based Set Decoding). The core innovation is separating the generative transport from atom-wise correspondence.

Architecture Overview

GLASS consists of three main components:

Permutation-Invariant Autoencoder:
- Encoder: A Transformer encoder processes atom tokens (types and fractional coordinates) and lattice embeddings. It uses attention pooling with a global learned query to map the unordered set of atoms into a fixed-dimensional, permutation-invariant global latent vector $z$ .
- Decoder: A "learned-slot" decoder reconstructs the structure. It utilizes $M$ learned slot embeddings (queries) that are concatenated with the global latent $z$ . These slots are processed through Transformer layers and decoded in parallel by separate heads for atom types, coordinates, and lattice parameters.
- Training: The autoencoder is trained to reconstruct inputs. Since the global bottleneck discards input ordering, the predicted atoms are matched to target atoms using the Hungarian algorithm to minimize a periodic squared Cartesian error and type mismatch cost. An auxiliary short-range pair-distance loss is added to improve local geometry.
Latent Flow Matching:
- Once the autoencoder is trained and frozen, a flow matching model is trained in the fixed-dimensional latent space.
- The model learns a vector field to transport Gaussian noise to the distribution of encoded crystal latents.
- Crucially, the flow operates entirely in the permutation-invariant latent space, completely bypassing the atom-wise correspondence problem during generation.
Generation Process:
- New latents are generated by integrating the learned flow from Gaussian noise.
- The global latent is decoded into a full all-atom structure in a single parallel pass using the learned slots.
- The slot identities provide a stable symmetry-breaking signal, allowing the model to produce distinct atoms without relying on a canonical ordering of the input.

3. Key Contributions

The paper outlines three primary contributions:

Identification of the Correspondence Problem as a Scaling Bottleneck: The authors isolate a failure mode where index-free permutation-equivariant particle flows require significantly more training as set size and density increase. They demonstrate that breaking symmetry with learned slots allows targets to be learned easily at every size.
Removal of Correspondence from Generative Transport: GLASS decouples the correspondence problem from the generative transport by operating flow matching in a permutation-invariant global latent space. Atom-wise correspondence is restricted to the reconstruction objective of the slot decoder, removing the need for imposed canonical atom ordering.
High-Validity Generation of Large MOFs: The model generates unconditional all-atom MOFs with up to 150 atoms per unit cell. It achieves structural validity approaching that of the training data without conditioning on building blocks, topology, or composition.

4. Experimental Results

MP20 Dataset (Small Crystals, ≤20 atoms)

Validity: After 1M flow training steps, GLASS reaches 98.38% raw validity, nearly matching the training data (98.45%) across all size bins. In contrast, other models (e.g., Zatom-1) show significant validity drops as atom count increases.
Geometry: Radial distribution functions of GLASS generations closely match training data peaks, whereas comparison models show distorted peaks or unphysical close contacts.
Tradeoff: A validity–novelty tradeoff is observed. The 50k checkpoint offers a balance (96.56% validity after relaxation, 52.71% novelty), while the 1M checkpoint maximizes validity but reduces novelty, with many generated structures recovering training samples.

QMOF150 Dataset (Large MOFs, ≤150 atoms)

Scalability: GLASS generates structures up to 150 atoms per cell. It achieves 74.8% raw structural validity (MOFChecker), significantly outperforming baselines like Mofasa (52.9%) and Zatom-1 (0.8%).
Validity vs. Novelty: While validity is high, novelty is limited. Under StructureMatcher analysis, 98.3% of valid GLASS samples match training structures. The authors attribute this to the autoencoder's generalization gap (held-out reconstruction RMSD of 1.25 Å vs. 0.02 Å on training data), suggesting the model struggles to decode unseen frameworks reliably, favoring familiar ones.
Autoencoder Gap: Experiments on the LeMat-Bulk dataset (5M structures, ≤20 atoms) show that with sufficient data, the autoencoder validation gap can be nearly closed, suggesting that the limitation on QMOF150 is partly due to dataset size and diversity rather than the architecture itself.

5. Significance and Claims

The paper claims that separating correspondence assignment from generative transport provides a simple and effective route toward high-validity generation of larger atomistic systems.

Mechanism: The fixed-grid experiments confirm that learned slot identities resolve the scaling difficulties inherent in index-free flows, allowing reliable generation at a constant training budget regardless of set size.
Limitations: The authors are modest regarding novelty. They acknowledge that while GLASS achieves near-training validity for large MOFs, most generated structures are reconstructions of known frameworks. The primary limitation identified is the autoencoder's ability to generalize to unseen structures, particularly in the context of the available QMOF150 data.
Future Directions: The work suggests that pre-training on larger MOF datasets could improve generalization and that global latent representations enable future directions such as property-informed latent training.

In summary, GLASS demonstrates that moving generative transport to a permutation-invariant latent space, combined with a slot-based decoder, overcomes the correspondence bottleneck that hinders the scaling of all-atom crystal generation, achieving high structural validity for systems up to 150 atoms.