Predicting how a molecule will pack itself into a solid-state crystal lattice is one of the most stubborn challenges in materials science and pharmaceutical development. A single molecule can often arrange itself into multiple distinct crystal structures—a phenomenon known as polymorphism. Each polymorph can possess vastly different physical properties. In medicine, a change in crystal structure can alter how quickly a drug dissolves in the body, potentially rendering a life-saving compound useless or toxic. In electronics, the spatial alignment of molecules determines how efficiently a material conducts electricity.
Yet, simulating and predicting these structures computationally remains incredibly difficult. Traditional methods are computationally expensive, often requiring massive supercomputing resources to search through millions of potential arrangements. While generative artificial intelligence has emerged as a promising alternative, standard generative models struggle with the strict geometric rules of crystallography. They often fail because they attempt to predict the positions of all atoms in a crystal simultaneously, which inevitably violates the delicate symmetry constraints that define real-world crystals.
The hurdle of crystal symmetry
To understand why generative artificial intelligence has struggled in this domain, it helps to examine how crystals are structured. A crystal is not just a random cluster of molecules; it is a highly ordered, repeating three-dimensional pattern. The fundamental building block of this pattern is the unit cell, which repeats infinitely in all directions. Within this unit cell lies the asymmetric unit, which is the smallest unique arrangement of atoms or molecules that contains no internal symmetry.
By applying specific mathematical rules of rotation, reflection, and translation, known as a space group, this asymmetric unit is copied and positioned to fill the entire unit cell. There are 230 unique space groups in nature, categorised into seven distinct crystal systems. For a crystal to remain stable and physically plausible, every single atom must conform precisely to these symmetry rules.
When conventional generative models attempt to design a crystal, they usually treat the entire unit cell as a collection of individual atoms. This is known as an atomistic full-cell approach. The model must learn not only how the atoms within a molecule bond together, but also how to place those atoms across the entire cell so that they perfectly align with the rules of the target space group. Even a tiny fraction of an Angstrom of misalignment can break the symmetry, resulting in an unstable structure that cannot exist in the real world. Furthermore, as the size of the system and the complexity of the symmetry group increase, the number of coordinates the model must track grows exponentially, making the generation process highly prone to errors.
Separating state and context
A new approach, detailed in a recent preprint that has not yet undergone peer review, attempts to solve this problem by rewriting the rules of how generative models interact with crystal geometry. The authors of the preprint have introduced a framework called SALA, which stands for Symmetry-Aware Flow Matching.
Instead of trying to generate the entire periodic crystal structure at once, SALA uses a technique called state-context separation. The core innovation lies in dividing the problem into two distinct parts: the state, which represents the active variables the model is trying to evolve, and the context, which represents the surrounding environment that influences those variables.
Specifically, SALA only evolves the coordinates of the atoms within the asymmetric unit, alongside the lattice parameters that are compatible with the target symmetry. Because the model is only manipulating the asymmetric unit, it is mathematically impossible for it to violate the rules of the space group. The symmetry is hard-coded into the generation process itself.
However, molecules do not exist in isolation; their arrangement is heavily dictated by how they interact with their neighbours across the crystal boundaries. To account for this, SALA uses a periodic context window. While the model only updates the coordinates of the asymmetric unit, it predicts these updates by analysing how the asymmetric unit fits into the full, repeating periodic environment. This allows the model to assess the intermolecular forces, electrostatic interactions, and physical boundaries from neighbouring cells without having to manually calculate or update those neighbouring coordinates.
By using a generative method known as flow matching—which establishes a smooth mathematical pathway to transition random noise into structured coordinates—SALA can generate molecular conformations, lattice geometries, and packing arrangements simultaneously in an end-to-end fashion.
Testing the limits of symmetry-aware generation
To evaluate the effectiveness of this approach, the authors tested the model against a dataset of 799 held-out crystal structures. This test set was highly diverse, spanning four distinct chemical classes and covering all seven crystal systems, ensuring that the model was not simply memorising a specific type of molecular arrangement.
When tasked with generating target crystal structures from molecular graphs and specified space groups, SALA demonstrated a significant performance leap over traditional methods. Generating 50 candidate structures per target, SALA achieved a 72.0 per cent hit rate. In comparison, a standard atomistic full-cell generative baseline managed a hit rate of just 8.1 per cent under the same conditions.
The model also showed a dramatic improvement in geometric accuracy. The average root-mean-square deviation for the best-candidate packing fell from 4.48 Angstroms for the baseline model to just 1.70 Angstroms for SALA. Notably, this advantage became even more pronounced as the systems grew larger and the symmetry operations became more complex.
Despite these promising results, it is important to recognise the boundaries of what this research proves. Because the study is currently published as a preprint, its findings have not yet been validated by independent peer review. Furthermore, the model is conditioned on both a molecular graph and a pre-specified space group. This means that while SALA is highly capable of finding the correct crystal packing once a target space group is chosen, it does not solve the broader challenge of predicting which space group a completely novel molecule will naturally prefer to crystallise in.
Additionally, while a 1.70 Angstrom deviation is a substantial improvement, it still represents a structural variance that could correspond to different energy states in real-world applications. The model serves as an efficient generator of high-quality candidates, but it is not yet a replacement for downstream quantum-chemical validation or experimental synthesis. It remains a computational tool positioned at the early-to-mid stages of the R&D pipeline.
Sources
- Symmetry-Aware Flow Matching for End-to-end Molecular Crystal Generation, Wendi Cai, Fanyang Mo, arXiv preprint, not yet peer reviewed, 2026-09-28
The R&D takeaway
For directors and funders of materials and pharmaceutical R&D, this work highlights the strategic value of hard-coding physical constraints directly into machine learning architectures rather than expecting models to learn them from raw data. By separating the active simulation state from its periodic context, teams can bypass the exponential computational scaling that typically hinders crystal structure prediction. This approach suggests that future investments in AI-driven materials discovery should focus on hybrid models that combine structural symmetry priors with flexible environmental context.
The R&D Innovate desk