2026 · KDD 2026

How Far Can You Grow? Characterizing the Extrapolation Frontier of Graph Generative Models for Materials Science

Can Polat , Erchin Serpedin , Mustafa Kurban , Hasan Kurban

Research highlights

RADII: locating the size limit of a materials generator

Measure when generated nanoparticles stop remaining reliable as requested size increases.

Why we investigated it

Performance within training sizes does not locate a model's extrapolation frontier.

What the work contributes

A radius-resolved benchmark isolates geometric extrapolation and measures how different structural errors emerge with size.

What the evidence shows

Five architectures fail in different sequences. Well-behaved models follow an approximately one-third scaling exponent, while scaling MatterGen stabilizes sampling without moving the observed frontier.

Build on this work

Use RADII's size splits and diagnostics to compare generators at and beyond their training radii.

  • Materials generation
  • Extrapolation frontier
  • Nanoparticles
  • Scaling laws
Abstract
Every generative model for crystalline materials harbors a critical structure size beyond which its outputs become unreliable; we call this the extrapolation frontier. Despite its consequences for nanomaterial design, this frontier has never been systematically measured. We introduce RADII, a radius-resolved benchmark of ~75,000 crystal-derived nanoparticle structures (33-11,298 atoms) that treats radius as a continuous scaling knob, tracing generation quality from in- to out-of-distribution under leakage-free splits. Each model is conditioned on target composition and atom count, isolating geometric extrapolation as the evaluation variable. RADII provides frontier-specific diagnostics: per-radius error profiles pinpoint each architecture's scaling ceiling, surface-interior decomposition separates boundary from bulk failures, and cross-metric sequencing reveals which aspect of structural fidelity breaks first. Benchmarking five state-of-the-art architectures, we find that: (i) well-behaved models degrade by ~13% in global positional error beyond training radii, while divergent models show poor fidelity across scales, with local bond fidelity ranging from negligible degradation to over 2x error growth; (ii) no two architectures share a failure sequence, revealing the frontier as a multi-dimensional surface shaped by model family; and (iii) well-behaved models follow the expected geometric scaling exponent alpha ~ 1/3, whose in-distribution fit predicts out-of-distribution error, making frontiers forecastable. Scaling MatterGen to its published parameter count stabilizes sampling but does not close the frontier, while DiffCSP remains unstable at published scale. These findings establish output scale as a first-class evaluation axis for geometric generative models. Code and data: https://github.com/KurbanIntelligenceLab/RADII.
Cite
@article{polat2026how,
  title = {How Far Can You Grow? Characterizing the Extrapolation Frontier of Graph Generative Models for Materials Science},
  author = {Can Polat and Erchin Serpedin and Mustafa Kurban and Hasan Kurban},
  year = 2026,
  journal = {KDD 2026},
  doi = {10.1145/3770855.3818872},
}