Original title: Symmetry- and Property-Aware Crystal Generation with Reinforcement Learning for Inverse Materials Design
Ting-Wei Hsu, Arun Bansil, Qimin Yan · September 2026
Designing new crystals with target physical traits is hard. For many crystal behaviors, a good score means nothing if the crystal structure lacks the right internal symmetry, which is balanced geometric order. Without that symmetry, the trait can be an illusion, fragile, or ill defined. To fix this, researchers created SPARC. It is a reinforcement learning tool that hunts for top properties while actively keeping the atomic symmetry needed to support them. They tested SPARC on two problems. First, it designed crystals with strong directional electrical responses, which only work in specific symmetry classes. Second, it maximized a solar device efficiency score, finding good crystal patterns on its own. It proves symmetry is essential for generating real, robust materials.
Suppose you measure the length of a shadow. The number you get depends entirely on where the sun is in the sky. If you do not know the sun's position, that measurement tells you nothing useful. In this paper, the crystal's symmetry is like the sun's position: without it, a property like how the crystal responds to electricity is not meaningful. The framework makes sure the symmetry is right before trusting the property value.
Materials scientists designing new crystals often optimize for a property like how a material responds to an electric field. But that response is only physically meaningful if the crystal's atomic arrangement has the right symmetry; otherwise the measured value could be an artifact. This framework ensures that generated crystals not only achieve the target property but also have the structural foundation to make that property real and usable, which is essential for creating reliable new materials for electronics and photonics.
The abstract demonstrates SPARC on only two specific tasks: uniaxial dielectric anisotropy and spectroscopic limited maximum efficiency. It does not discuss how the framework scales to more complex properties or larger crystal structures. Additionally, for properties that do not have a prescribed symmetry class, the framework may need to explore a wider space, and the abstract does not address computational efficiency or convergence guarantees.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Symmetry- and Property-Aware Crystal Generation with Reinforcement Learning for Inverse Materials Design — Ting-Wei Hsu, Arun Bansil, Qimin Yan
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.