StructEval: A Benchmark for Evaluating Generation, Inference, and Reconstruction in Atomic and Crystalline Structures

ICLR 2026 Conference Submission20754 Authors

Keywords: dataset, benchmark, crystal materials, generation, quantum chemistry, computational materials

Abstract: Crystalline materials power technologies from solar conversion to catalysis, yet current machine learning evaluations artificially divide infinite lattices and finite nanoclusters into separate domains. StructEval unifies these regimes with a symmetry-aware, radius-resolved framework that systematically links primitive unit cells to nanoparticles across ten industrially critical compounds. Each material includes 20+ radii configurations between 0.6-3.0 nm, sampled at 780 quasi-uniform orientations, producing nearly $200,000$ structures spanning $55-11,298$ atoms. StructEval defines three rigorous challenges--unit cell $\rightarrow$ nanoparticle generation, nanoparticle $\rightarrow$ unit cell resolution, and lattice reconstruction--supported by a leakage-free split that isolates orientation interpolation from radius extrapolation. This design enables precise measurement of generalization across scale and symmetry. Benchmarking leading generative models reveals severe breakdowns under out-of-distribution conditions, exposing a fundamental gap in current architectures. By providing a reproducible, geometry-grounded testbed fully compatible with PyTorch Geometric, StructEval establishes the foundation for next-generation generative, inference, and reconstruction models in crystalline systems. Data and implementations are released at https://anonymous.4open.science/r/StructEval-ANONYMOUS.

Primary Area: datasets and benchmarks

Submission Number: 20754