2026 · Zenodo (CERN European Organization for Nuclear Research)
SCALAR: Cross-Scale Benchmark of Nanoparticle Structures for Quantifying Hallucination, Consistency, and Reasoning in Materials Foundation Models
Can Polat , Erchin Serpedin , Mustafa Kurban , Hasan KurbanResource highlights
SCALAR seed data for evaluating materials foundation models
Construct controlled size and rotation shifts for materials-model evaluation.
Why we investigated it
Testing hallucination across scale requires known structures with consistent provenance and reproducible transformations.
What the work contributes
The archive provides primitive cells and base nanoparticles for 83 materials; a separate pipeline creates the augmented benchmark.
What the evidence shows
The versioned record documents a correction to one NaRhH3 structure and identifies which remaining structures are unchanged.
Build on this work
Use the corrected archive and generation pipeline, recording the dataset version when comparing model results.
- SCALAR dataset
- Structural hallucination
- Nanoparticles
- Scale generalization
Abstract
SCALAR (Structural Consistency And Logic Across Regimes) is a cross-scale benchmark of ≈100,000 nanoparticle structures (25–18,123 atoms across 83 predominantly hydride materials, carving radii 10–30 Å) designed to evaluate structural hallucination, cross-scale consistency, and physical reasoning in materials foundation models. Each structure is a spherical truncation of a DFT-relaxed, experimentally validated crystal lattice; the benchmark treats radius as a controlled scaling knob and provides leakage-free in-distribution and out-of-distribution splits over both radius and SO(3) rotation. This archive contains the seed inputs for the benchmark: the 83 primitive unit cell CIFs and per-(material, radius) base XYZ structures. The full benchmark (with rotation augmentation) is reproduced locally by running the generation pipeline shipped in the GitHub repository. Paper: "SCALAR: Quantifying Structural Hallucination, Consistency, and Reasoning Gaps in Materials Foundation Models" (submitted to Digital Discovery). Code & generation pipeline: github.com/KurbanIntelligenceLab/SCALAR (MIT license). To reproduce the full benchmark from this archive: python -m create_scalar.create_scalar --raw-data scalar_raw.zip --output scalar Version 1.1 (correction). Replaces scalar_raw/NaRhH3/NaRhH3_R30.xyz, which in version 1.0 contained 11,917 atoms — identical to the R = 29 Å structure and the only structure in the archive whose atom count did not increase with radius. Its effective density was 10.2% below the value stable across R = 10–29 Å for this material. The corrected structure, regenerated with create_scalar/carve.py, contains 13,249 atoms at the expected density. All other 1,742 structures are byte-identical to version 1.0.
Cite
@article{polat2026dataset,
title = {SCALAR: Cross-Scale Benchmark of Nanoparticle Structures for Quantifying Hallucination, Consistency, and Reasoning in Materials Foundation Models},
author = {Can Polat and Erchin Serpedin and Mustafa Kurban and Hasan Kurban},
year = 2026,
journal = {Zenodo (CERN European Organization for Nuclear Research)},
doi = {10.5281/zenodo.20631921},
}