cs.LG

Building crystals by breaking symmetry — a diffusion model that mirrors a physical phenomenon

cs.LG Van Khoa Nguyen, Alexandros Kalousis Aug 2026

Generative models for crystals have not produced complete crystallographic specifications, sampling space groups from empirical distributions instead. This work, drawing on spontaneous symmetry breaking in physics, reverses from the lowest-symmetry priors.

Paper overview (our summary)

  • Field (arXiv category)cs.LG
  • AuthorsVan Khoa Nguyen, Alexandros Kalousis
  • Submitted2026-08-13
  • arXiv ID2608.13457v1

Key points

  • SbCD is a diffusion-based framework generating complete structure specifications for crystals.
  • State-of-the-art approaches generate only as far as site symmetries and sample space groups from empirical distributions.
  • Drawing on spontaneous symmetry breaking in physics, it reverses from the lowest-symmetry priors.
  • A Markovian jump-diffusion process represents the symmetry-breaking dynamics and traverses space groups in a physically motivated way.
  • On de novo generation over MP20 and MPTS-52 it is reported to beat its symmetry-preserving counterpart by a substantial margin.

1Mirroring a phenomenon in a procedure

The previous article followed a design that supplied structure from the subject side, disease context, as a condition. What this work mirrors is the phenomenon itself. The physical behaviour by which a crystal breaks symmetry under external conditions is carried over into the procedure that generates one.

2What could not be produced before

The aspectApproaches so farThis work, SbCD
How far generation reachesAs far as site symmetriesA complete structure specification
Treatment of space groupsSampled from an empirical distributionHandled as transitions within the generative process
Starting pointThe lowest-symmetry priors
Character of the processA Markovian jump-diffusion representing symmetry breaking

Rather than handing the space group in from outside, it is treated as something that shifts during generation. Since symmetry breaking is an event on the physics side, the reasoning goes, it belongs inside the process rather than in a sampling step.

3Reversing, with a meaning

  1. 1What happens in physicsA crystal breaks symmetry under external conditions
  2. 2The counterpart in generationBegin from the lowest-symmetry priors
  3. 3Design of the processA Markovian jump-diffusion carries transitions between space groups
  4. 4What comes outA full crystallographic specification, not one stopping at site symmetries

Diffusion models generally reverse from noise toward structure. Here that reversal is given a physical meaning: from lower symmetry toward higher.

4The evaluation

De novo generation was run on MP20 and MPTS-52, where the method is reported to beat its symmetry-preserving counterpart by a substantial margin. What matters is what it was compared against. Holding the framework fixed and changing only the treatment of symmetry is a comparison built so that the effect of building in the breaking can be seen.

Why it matters

When a generative model is brought into a natural science, how much of the structure on the subject side gets mirrored in the procedure becomes the design question. Whether a space group is sampled in from outside or treated as something that shifts during generation is one instance. Mirroring a physical phenomenon in the process itself changes what the model represents before it changes how well it performs.

FAQ

What is a space group?
A classification of the symmetry a crystal has. The paper treats sampling it from an empirical distribution as a limitation of existing methods.
Why start from the lowest-symmetry priors?
To carry into the generative process the physical behaviour by which crystals break symmetry under external conditions.
What was it compared against?
Its symmetry-preserving counterpart, so that only the treatment of symmetry differs between the two.

Sources (primary)

Source: arXiv (descriptive metadata is CC0 public domain). Summaries are our own; see arXiv for the original text and PDF.

#AI research#Materials science#Generative models#arXiv#Physics
Disclaimer: This site independently summarizes and classifies information based on official data sources. Always verify the latest and accurate information with the official sources. Content on finance, health, legal, and security is information, not advice. This site is not an official website of the U.S. government.