🤖 machine learning

Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion

The paper introduces Symmetry-breaking Crystal Diffusion (SbCD), a novel diffusion-based framework that utilizes Markovian jump processes to spontaneously break symmetries and generate complete crystal structures with global structural dependencies, significantly outperforming existing symmetry-preserving models on standard benchmarks.

Original authors: Van Khoa Nguyen, Alexandros Kalousis

Published 2026-08-14

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Van Khoa Nguyen, Alexandros Kalousis

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

Imagine the universe is built from tiny, invisible LEGO bricks called atoms. When these atoms snap together in a repeating, orderly pattern, they form crystals. These aren't just pretty rocks; they are the secret sauce behind everything from the batteries in your phone to the chips in your computer and even the medicines that keep us healthy. For decades, scientists have tried to design new, better crystals by mixing and matching these atomic bricks, but it's been a slow, frustrating game of trial and error, often taking years to find a single useful structure.

In recent years, computers have started helping out using "generative models." Think of these as digital chefs that learn to cook up new crystal recipes by studying millions of existing ones. However, most of these digital chefs have a major blind spot: they are told to stick to a specific, rigid rulebook (called a "space group") before they even start cooking. They can't imagine breaking the rules to create something entirely new. In the real world, crystals often change their internal rules as conditions shift, a phenomenon known as "symmetry breaking." Until now, our digital chefs haven't been able to mimic this natural flexibility, limiting their ability to discover truly novel materials.

Enter a new approach called Symmetry-breaking Crystal Diffusion (SbCD), a method that teaches the computer to be a bit more rebellious. Instead of forcing the crystal to follow a strict rulebook from the start, SbCD starts with the most chaotic, rule-free version of a crystal possible (the lowest symmetry). Then, like a sculptor revealing a statue from a block of marble, it gradually reverses the process of symmetry breaking, allowing the crystal to evolve from a simple, low-symmetry state back into a complex, ordered structure. This process is inspired by how real crystals behave in nature, where they often shift from high-symmetry states to lower-symmetry ones under pressure or heat.

The researchers behind this method, Van Khoa Nguyen and Alexandros Kalousis, used a mathematical tool called "Markovian jump diffusion" to model these shifts. Imagine a game where a crystal starts in a high-symmetry state and, at random moments, "jumps" down to a simpler, lower-symmetry state. The computer learns to play this game in reverse: starting from the simplest, most broken state (the lowest symmetry) and "jumping" back up to create complex, stable crystals. By doing this, the model doesn't just copy existing patterns; it learns the underlying logic of how crystals form and change.

When they tested SbCD on two major crystal databases (MP-20 and MPTS-52), the results were promising. The model generated new crystal structures that were not only valid and chemically sound but also more likely to be thermodynamically stable (meaning they wouldn't fall apart) compared to previous methods. Crucially, SbCD was the first to successfully generate complete crystal specifications—including the specific symmetry rules and atomic positions—without needing to be told which rules to follow beforehand. While the paper suggests this is a significant step forward for in silico (computer-based) material discovery, the authors note that this is a proof of concept. The model still needs to be tested in real-world labs, and future work will focus on making the "jumps" between symmetry states even more realistic to reflect the complex physics of actual materials. But for now, this new digital sculptor has shown that letting a crystal break its own rules might be the key to building the next generation of technology.

Technical Summary: Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion

Problem Statement

The discovery of new crystalline materials is critical for advancements in energy storage, semiconductors, and pharmaceuticals, yet traditional experimental discovery is slow and costly. While generative models have emerged as a tool for in silico screening, existing approaches face significant limitations in capturing global symmetry and structural dependencies. Current state-of-the-art methods typically generate crystals only up to site symmetries or rely on sampling space groups from empirical distributions during generation. This reliance on pre-defined or empirically sampled space groups limits the models' ability to learn the full distribution of crystallographic structures and often results in poor generalization to unseen configurations, particularly regarding Wyckoff positions. Furthermore, many methods fail to explicitly model the inter-dependencies between space groups, lattice constraints, and site symmetries, treating them as separate or fixed constraints rather than dynamic components of the generative process.

Methodology

The authors propose Symmetry-breaking Crystal Diffusion (SbCD), a novel diffusion-based framework that generates complete crystallographic structure specifications by reversing from the lowest-symmetry priors. The methodology is inspired by spontaneous symmetry breaking (SSB) in physics, where ordered phases emerge through the breaking of higher symmetries under external conditions.

Core Framework

SbCD unifies diffusion-based training objectives across continuous and discrete time and state spaces. The model represents a crystal via an asymmetric unit tuple $M = (F, A, k, S, G)$ , where $F$ are fractional coordinates, $A$ are atom types, $k$ is the invariant lattice representation, $S$ is site symmetry, and $G$ is the space group.

Markovian Jump Diffusion for Space Groups:
Unlike previous methods that fix the space group, SbCD models the space group $G$ as a dynamic variable evolving via a Markovian jump-diffusion process.
- Forward Process: Simulates SSB by transitioning from a higher-symmetry space group $\tilde{G}$ to a lower-symmetry group $G$ (eventually reaching $P1$ , the triclinic system with only identity operations) at a random time step $\tau$ . This is governed by a time-dependent rate matrix $\Lambda_t$ .
- Reverse Process: Learns to transition from the lowest-symmetry prior ( $P1$ ) back to higher-symmetry space groups, effectively "recovering" the symmetry.
- Variational Bound: The authors derive a variational upper bound on the negative log-likelihood that decomposes into terms for atom types, fractional coordinates, space groups, and conditional lattice/site-symmetry terms.
Symmetry-Breaking Diffusion on Continuous State Space (Lattice):
The lattice representation $k$ is constrained by the crystal family and space group. SbCD introduces a symmetry-breaking diffusion process where lattice constraints adaptively change as the space group transitions.
- The forward kernel applies a masking mechanism that respects the constraints of the current space group. When a jump occurs at time $\tau$ , the mask and bias parameters update to reflect the new, lower-symmetry constraints.
- The reverse process utilizes an analytical form that interpolates between the constraints of the previous and current space groups, ensuring the generated lattice remains valid throughout the trajectory.
Symmetry-Breaking Diffusion on Discrete State Space (Site Symmetry):
To model site symmetries $S$ , the authors propose a simplified one-hot vector representation of 81 distinctive oriented site-symmetry symbols, replacing the complex binary matrix used in prior work (e.g., SymmCD).
- Interpolated Prior: As the space group transitions from $\tilde{G}$ to $G$ , the marginal prior over site symmetries shifts. The authors define a time-dependent prior $\pi_t$ that linearly interpolates between the marginal priors of the two space groups ( $\tilde{g}$ and $g$ ) along a straight line in the probability simplex.
- This allows the diffusion process to smoothly transition site-symmetry states as the global symmetry breaks, avoiding the need to project generated states onto the nearest valid point group.

Key Contributions

Theoretical Unification: The paper derives a variational bound that explicitly unifies the structural dependencies among crystal components, treating the space group distribution as a learnable component rather than a fixed constraint.
Markovian Jump Diffusion: It introduces a principled approach to model space-group transitions using Markovian jump diffusion, enabling the traversal of different space groups in a physically motivated manner (mimicking SSB).
Adaptive Constraints: The framework develops novel diffusion processes for both continuous (lattice) and discrete (site symmetry) state spaces that adaptively enforce space-group constraints during the diffusion trajectory.
Simplified Representation: A simplified one-hot representation for site symmetries is proposed, which facilitates capturing site-symmetry distributions and yields more stable generated structures compared to previous binary encodings.
Complete Specification Generation: SbCD is presented as the first generative framework capable of producing complete crystallographic structure specifications, including space groups and Wyckoff positions, without relying on empirical sampling of these elements during inference.

Experimental Results

The authors evaluate SbCD on the MP-20 and MPTS-52 datasets, comparing it against baselines including CDVAE, DiffCSP, FlowMM, and the symmetry-preserving counterpart SymmCD.

MP-20 Performance: SbCD variants (specifically $SbCD^\diamond$ ) outperform symmetry-preserving counterparts and other baselines in generating thermodynamically stable structures. They show significantly higher proportions of charge-neutral compositions and better coverage of the test set's compositional distribution (measured by unique elements).
Stability and Validity: While some baselines achieve high structural validity, they often generate low-density, loosely packed crystals. SbCD variants achieve a better trade-off, generating structurally valid crystals with mean density ratios closer to the training data.
Space Group Distribution: SbCD narrows the performance gap with methods conditioned on empirical space-group distributions, outperforming methods that do not explicitly model symmetries. It demonstrates the ability to capture a wide range of space-group distributions.
Scalability: On the more challenging MPTS-52 dataset (up to 52 atoms), SbCD scales effectively (tested with -S, -L, and -XL model sizes) and achieves the best performance on thermodynamic stability metrics among the compared methods.
Efficiency: In low-number-of-function-evaluations (NFE) regimes, SbCD significantly outperforms SymmCD in generating valid structures, suggesting potential for efficient few-step generation.

Significance and Claims

The paper claims that SbCD represents a significant step forward in de novo crystal generation by moving beyond the limitations of symmetry-preserving models and empirical space-group sampling. By explicitly learning the transition dynamics between space groups and adapting constraints accordingly, the model captures global symmetry and structural dependencies that previous methods missed.

The authors position SbCD as a "proof of concept" for generative modeling of crystalline materials that can produce complete crystallographic specifications from minimal symmetry assumptions. They note that while the current framework uses direct transitions to $P1$ , future work could explore more realistic physical transition paths. The method is presented as a tool to accelerate materials discovery by efficiently exploring large chemical and crystallographic design spaces, though the authors acknowledge the need for responsible use and rigorous experimental validation for any generated materials.