PEROV-H3: Evaluating Generative Models under Size and Symmetry Shifts in Hydrogen-Storage Perovskites

Can Polat, Erchin Serpedin, Mustafa Kurban, HASAN KURBAN

Keywords: crystal generation, dataset, hydrogen storage, perovskites, generative ai

Abstract: We introduce PEROV-H3, a rigorous benchmark targeting ABH$_3$ perovskites, designed to evaluate generative models under controlled size and symmetry shifts with structure-aware metrics. In materials science, models often excel on ideal, periodic crystals yet degrade on finite nanoparticles where size, surfaces, and edges dominate. PEROV-H3 closes this gap by pairing two complementary tasks: $(i)$ unit-cell $\rightarrow$ nanoparticle generation, probing surface- and size-dependent distortions; and $(ii)$ nanoparticle $\rightarrow$ unit-cell reconstruction, recovering bulk lattice parameters and symmetry. The benchmark comprises 100 DFT-relaxed ABH$_3$ compositions and 210,000 nanoparticle configurations spanning radii $R\in\{6,\dots,30\}$ Angstrom (systematic size splits for ID/OOD). Baselines reveal substantial errors under extrapolation, especially in symmetry and lattice recovery, indicating that current models memorize templates rather than learn the physics of scale. PEROV-H3 thus provides a chemically diverse, size-systematic, and structurally clean testbed for stress-testing generative models beyond bulk crystals. The dataset and the implementation are available at https://anonymous.4open.science/r/PEROV-H3.

Primary Area: datasets and benchmarks

Submission Number: 8567