Symmetry- and Property-Aware Crystal Generation with Reinforcement Learning for Inverse Materials Design

Technical Summary: Symmetry- and Property-Aware Crystal Generation with Reinforcement Learning (SPARC)

Problem Statement
The inverse design of crystalline materials aims to identify structures with specific physical properties. However, many functional responses—such as dielectric anisotropy, nonlinear optical effects, and symmetry-protected electronic phenomena—are not determined solely by composition but are fundamentally constrained by the symmetry of the underlying crystal lattice. Existing generative models, particularly those based on diffusion or continuous coordinate representations (e.g., latent diffusion models), often optimize for scalar objectives (like formation energy or band gap) without explicitly preserving crystallographic symmetry. Consequently, these models may generate structures that appear to satisfy a target property numerically but lack the necessary symmetry to make that property physically robust or well-defined. For instance, a generated crystal might exhibit accidental dielectric anisotropy that collapses under minor structural distortions or fails to exist in the required point group. There is a critical need for generative frameworks that optimize functional objectives while strictly adhering to the symmetry conditions under which those objectives are physically admissible.

Methodology: The SPARC Framework
The authors introduce SPARC (Symmetry- and Property-Aware Reinforcement Learning for Crystal Generation), a framework that couples a symmetry-constrained diffusion generator with reinforcement learning (RL) to optimize physical objectives while preserving structural symmetry.

1. Symmetry-Constrained Generation Backbone:
   SPARC utilizes SymmCD, a diffusion model that generates crystals in the asymmetric-unit representation. Instead of diffusing atoms in continuous 3D space, the model samples a space group $G$ and the number of representative sites $M$. It then denoises the asymmetric unit $M' = (A', F', \Sigma', k)$, comprising atom types, fractional coordinates, site-symmetry labels, and a symmetry-constrained lattice basis. The full crystal is reconstructed by applying the operations of $G$ to these representatives. This ensures that every generated structure inherently possesses the symmetry of the sampled space group.

2. Reinforcement Learning Loop:
   The framework employs a two-loop RL strategy to steer generation toward target properties:

   • Inner Loop (Policy Optimization): The diffusion policy $p_\theta$ is fine-tuned using a reward-weighted policy gradient. The objective maximizes the expected terminal reward $r(M)$ while maintaining proximity to the pre-trained generator $p_{pre}$ via a Kullback–Leibler (KL) divergence penalty. This prevents "reward hacking" and preserves the broad chemical and structural coverage learned during pretraining.
   • Outer Loop (Space-Group Proposal): Since the diffusion policy conditions on a fixed space group $G$ and cannot reshape the distribution of $G$ itself, SPARC implements an adaptive outer loop. It updates the proposal distribution $\pi^{(j)}(G)$ based on the accumulated rewards of realized space groups. This allows the framework to shift sampling probability toward symmetry classes that support the target property while maintaining a "probability floor" to prevent premature collapse onto a single low-symmetry group.

3. Reward Evaluation:
   Generated structures are relaxed using a machine-learning interatomic potential (MatterSim) and filtered for stability, uniqueness, and novelty (S.U.N.). Target properties are evaluated using surrogate models (modified Tensorial Spectra Equivariant Neural Networks, TSENN) to assign terminal rewards.

Key Results
The authors evaluate SPARC on two complementary inverse-design tasks:

1. Uniaxial Dielectric Anisotropy (Symmetry-Sensitive Task):

   • Objective: Maximize the contrast between in-plane ($\varepsilon_{\parallel}$) and out-of-plane ($\varepsilon_{\perp}$) dielectric responses, requiring $\varepsilon_{xx} = \varepsilon_{yy} \neq \varepsilon_{zz}$. This tensor form is only symmetry-enforced in trigonal, tetragonal, and hexagonal crystal classes.
   • Outcome: Over 120 RL steps, the mean reward increased from 0.17 to 0.56. The space-group distribution shifted significantly, with trigonal, tetragonal, and hexagonal groups (e.g., $P\bar{3}1m$, $P4_2/mnm$) comprising ~44% of generated structures.
   • Ablation Study: Comparing SymmCD (symmetry-aware) against DiffCSP (symmetry-agnostic) revealed a stark difference. While both achieved similar numerical rewards, DiffCSP generated structures that were predominantly $P1$ or $P\bar{1}$ (90% without relaxation), relying on accidental anisotropy. SymmCD produced structures with enforced high-order rotational axes (3-fold, 4-fold, 6-fold) in 50.1% of cases, yielding a 35× higher efficiency in finding symmetry-protected candidates.
   • Validation: First-principles DFT validation confirmed that candidates like Blue Phosphorus ($P3_1$) and $ZrTi_2O_6$ ($P\bar{3}1m$) retained their symmetry-enforced uniaxial anisotropy after relaxation, with Blue Phosphorus showing a ratio $\varepsilon_{\parallel}/\varepsilon_{\perp} \approx 3.46$.

2. Spectroscopic Limited Maximum Efficiency (SLME) (Symmetry-Agnostic Task):

   • Objective: Maximize the SLME ($\eta$) of single-junction photovoltaic absorbers, a scalar objective dependent on the full frequency-resolved optical response and band gap, without a prescribed symmetry class.
   • Outcome: Fine-tuning shifted the population from low-gap/low-efficiency regions to a target region ($E_g \approx 0.8\text{--}1.9$ eV, $\eta \ge 25\%$). The fraction of structures meeting these criteria increased from 11% to 32%.
   • Diversity: Unlike the dielectric task, the SLME task did not collapse the symmetry distribution. The cubic $Fm\bar{3}m$ group remained dominant, but orthorhombic ($Pnma$) and triclinic ($P\bar{1}$) groups also increased in prevalence. The chemical diversity was preserved (46 elements represented), with enrichment in sulfur and antimony, consistent with known photovoltaic materials.

Significance and Claims
The paper claims that SPARC demonstrates that symmetry and functional properties must be optimized together rather than as independent stages.

• Symmetry as a Physical Foundation: The authors argue that symmetry is not merely an additional constraint but a physical foundation. Without it, generated responses may be ill-defined, accidental, or unstable.
• Necessity of Symmetry-Aware Backbones: The ablation study explicitly shows that reward optimization alone cannot guarantee symmetry-protected properties; a symmetry-aware generation backbone is required to enforce the structural conditions necessary for the physics to exist.
• Adaptive Discovery: SPARC can identify favorable symmetry classes when they are not known in advance (as in the SLME task) or enforce specific classes when required (as in the dielectric task).
• Robustness: By generating candidates that satisfy both electronic/optical constraints and crystallographic plausibility, SPARC produces realistic materials candidates that are more likely to be experimentally viable than those from standard generative models.

The work establishes a route toward inverse materials design where the search is guided not only toward high-performing materials but toward structures that can physically support the targeted response through their inherent symmetry.