C2NP: A Benchmark for Learning Scale-Dependent Geometric Invariances in 3D Materials Generation
Can Polat Texas A&M University [email protected] Erchin Serpedin Texas A&M University [email protected] Mustafa Kurban * Ankara University Texas A&M University at Qatar [email protected] Hasan Kurban * Hamad Bin Khalifa University [email protected] · 2026
Read C2NP: A Benchmark for Learning Scale-Dependent Geometric Invariances in 3D Materials Generation by Can Polat Texas A&M University [email protected] Erchin Serpedin Texas A&M University [email protected] Mustafa Kurban * Ankara University Texas A&M University at Qatar [email protected] Hasan Kurban * Hamad Bin Khalifa University [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-19076.txt.
Introduction
Can Polat and four co-authors test whether models that dream up materials can jump across scales. Models for making new materials do well on bulk crystals that repeat without end. Real use often needs finite grains, where surfaces rule and size warps the shape, and skill across that jump is mostly untested.
Their bench, C2NP, scores models as they move from unit cells to nanoparticles. It checks whether the deep patterns a model learned at one scale still hold at the other. A clear scale test matters because drugs, catalysts, and devices all live at the small end of the scale.
Catalog record
| CREATOR | Can Polat Texas A&M University [email protected] Erchin Serpedin Texas A&M University [email protected] Mustafa Kurban * Ankara University Texas A&M University at Qatar [email protected] Hasan Kurban * Hamad Bin Khalifa University [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-19076
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
C2NP: A Benchmark for Learning Scale-Dependent Geometric Invariances in 3D Materials Generation Can Polat Texas A&M University [email protected] Erchin Serpedin Texas A&M University [email protected] Mustafa Kurban * Ankara University Texas A&M University at Qatar [email protected] Hasan Kurban * Hamad Bin Khalifa University [email protected] Abstract Generative models for materials have achieved strong performance on periodic bulk crystals, yet their ability to generalize across scale transitions to finite nanostructures remains largely untested. We introduce Crystal-to-Nanoparticle (C2NP) , a systematic benchmark for evaluating generative models when moving between infinite crystalline unit cells and finite nanoparticles, where surface effects and size-dependent distortions dominate. C2NP defines two complementary tasks: (i) generating nanoparticles of specified radii from periodic unit cells, testing whether models capture surface truncation and geometric constraints; and (ii) recovering bulk lattice parameters and space-group symmetry from finite particle configurations, assessing whether models can infer underlying crystallographic order despite surface perturbations. Using diverse materials as a structurally consistent testbed, we construct over 170,000 nanoparticle configurations by carving particles from supercells derived from DFT-relaxed crystal unit cells, and introduce size-based splits that separate interpolation from extrapolation regimes. Experiments with state-of-the-art approaches, including diffusion, flow-matching, and variational models, show that even when losses are low, models often fail geometrically under distribution shift, yielding large lattice-recovery errors and near-zero joint accuracy on structure and symmetry. Overall, our results suggest that current methods rely on template memorization rather than scalable physical generalization. C2NP offers a controlled, reproducible framework for diagnosing these failures, with immediate applications to nanoparticle catalyst design, nanostructured hydrides for hydrogen storage, and materials discovery. Dataset and code are available at https://github.com/KurbanIntelligenceLab/C2NP . I. Structural Construction Unit cell \mathcal{U} \Lambda=(a,b,c,\alpha,\beta,\gamma) Supercell Tiling \mathcal{T}=20^{3} Nanoparticle \mathcal{P}_{R} Carve sphere of radius R around \mathbf{x}_{0} II. Orientational Augmentation Over SO(3) Training set \mathcal{Q}_{\text{train}} Sparse grid: \theta_{\text{train}}=15^{\circ} ( N\!\approx\!59 ) Deterministic seeding over SO(3) ID set \mathcal{Q}_{\text{ID}} Medium grid: \theta_{\text{ID}}=12^{\circ} ( N\!\approx\!92 ) Split margin: d\geq\delta_{\text{ID}}=6^{\circ} , offset R_{\text{ID}}q OOD set \mathcal{Q}_{\text{OOD}} Dense grid: \theta_{\text{OOD}}=9^{\circ} ( N\!\approx\!163 ) Split margin: d\geq\delta_{\text{OOD}}=4.5^{\circ} , offset R_{\text{OOD}}q III. Benchmark Tasks Task 1: Unit cell \rightarrow nanoparticle Input: (\mathcal{U},R) , Output: \mathcal{P}_{R} Metrics: RMSD, Hausdorff distance, hull volume… Task 2: Nanoparticle \rightarrow lattice Input: \mathcal{P}_{R} , Output: (\Lambda,\Gamma) Metrics: lattice RMSE, space-group accuracy, joint recovery Figure 1: C2NP data generation and evaluation pipeline. The first stage (green) constructs finite nanoparticles from periodic crystallographic inputs: starting from each DFT-optimized unit cell \mathcal{U} with lattice parameters \Lambda=(a,b,c,\alpha,\beta,\gamma) , a 20\times 20\times 20 supercell \mathcal{T} is built to ensure sufficient spatial extent, after which a spherical cluster \mathcal{P}_{R} is carved by retaining atoms within radius R of a chosen center \mathbf{x}_{0} . The second stage (blue) applies rotational augmentation over \mathrm{SO}(3) using a geodesic separation constraint on unit quaternions to generate three stratified orientation grids: a sparse training set \mathcal{Q}_{\mathrm{train}} with threshold \theta_{\mathrm{train}}=15^{\circ} , an interpolative in-distribution set \mathcal{Q}_{\mathrm{ID}} with \theta_{\mathrm{ID}}=12^{\circ} and exclusion margin \delta_{\mathrm{ID}}=6^{\circ} , and a dense extrapolative out-of-distribution set \mathcal{Q}_{\mathrm{OOD}} with \theta_{\mathrm{OOD}}=9^{\circ} and margin \delta_{\mathrm{OOD}}=4.5^{\circ} . Deterministic quaternion seeding and fixed offset rotations R_{\mathrm{ID}} and R_{\mathrm{OOD}} enforce non-overlapping geodesic neighborhoods between train, ID, and OOD orientations while decorrelating orientation manifolds. The third stage (orange) defines two complementary benchmark tasks: a forward problem mapping unit-cell parameters and target radius (\mathcal{U},R) to the corresponding nanoparticle geometry \mathcal{P}_{R} , and an inverse problem recovering lattice parameters and symmetry (\Lambda,\Gamma) from a nanoparticle \mathcal{P}_{R} , providing a unified testbed for both generative and inverse lattice–nanoparticle modeling. Figure 2: Representative crystal-to-nanoparticle transformations in the C2NP dataset, derived from DFT-validated crystallographic reference structures. Columns (left to right) correspond to ZrCdH 3 , LiBeH 3 , LiCaH 3 , BaMnH 3 , and LiBeH 3 . Rows show the reference unit cell (top), a nanoparticle carved at R=10 ,Å (middle), and a nanoparticle carved at R=30 ,Å (bottom). Each panel reports the total atom count ( N ) and elemental composition. Across the dataset, particle sizes span more than four orders of magnitude, from a minimum of N=4 atoms for the smallest particles to a maximum of N=18{,}123 atoms for the largest configurations. Atom colors follow the standard CPK scheme: H (white), Li (purple), Be (dark green), Ca (green), Ba (light green), Mn (purple), Zr (dark gray), and Cd (light gray). 1 Introduction Computational materials modeling generally falls into two distinct but interconnected domains. The first domain concerns infinite periodic crystals, characterized by unit cells that encode lattice constants, atomic basis vectors, and space group operations 37 . This crystallographic representation, standardized through CIF formats, distills the translational symmetry of bulk solids and forms the conceptual backbone of solid-state physics and quantum chemistry 20 . The second domain encompasses nanoparticles, finite atomic assemblies where translational periodicity is absent. Here, surface facets, edge sites, and reduced coordination numbers give rise to structural relaxations, electronic localization, and quantum size effects that profoundly modify functional properties 31 . Many technologically relevant materials exist at the intersection of these descriptions, and characterizing how bulk order transitions to finite-scale behavior is essential for understanding catalysis, optoelectronics, thermodynamic stability, and charge transport 39 . Although recent progress in machine learning has transformed materials discovery, connecting bulk and nanoscale regimes remains an open problem 26 . Neural models trained on crystallographic databases often generalize well within periodic systems but exhibit significant degradation when tasked with nanoparticle generation or lattice inference from finite configurations 11 . Common failure modes include incorrect space group assignment, erroneous lattice constant predictions, and poor representation of size-driven surface effects 14 . Current benchmarks expose these shortcomings: CSPBench demonstrates that leading crystal prediction methods collapse when evaluation distributions deviate from training data 40 . Meanwhile, application-focused perovskite datasets prioritize property prediction tasks, such as bandgap estimation or solar cell efficiency, without addressing bidirectional structural transformations or controlled size-dependent variation 25 . To address this gap, we present C2NP , a benchmark that systematically tests whether generative models can handle the structural transformations inherent to bulk-to-nanoparticle transitions. Our approach leverages perovskite hydrides, a material family with well-established relevance to energy storage and heterogeneous catalysis 23 , as a structurally coherent platform. We construct a dataset pairing DFT-validated unit cells with about 172,000 nanoparticle configurations derived by spherical carving across radii from 6 to 30 Å with 1 Å steps. This construction strategy isolates the impact of finite size and surface effects while maintaining chemical diversity, avoiding the confounding influence of disparate crystal topologies that would complicate interpretation 34 . The benchmark is organized around two complementary evaluation protocols: forward generation, where models must produce nanoparticles of specified radius from a given unit cell (requiring understanding of how periodic lattices terminate at finite boundaries), and inverse reconstruction, where models infer bulk lattice parameters and space group symmetry from nanoparticle structures despite surface disorder and coordinate ambiguity. This bidirectional design probes whether learned representations encode scalable structural principles or merely memorize correlations within the training distribution. C2NP thus provides not only a diagnostic benchmark for identifying generalization failures but also a foundation for developing architectures that can reason about physical scale in crystalline matter. Figure 1 outlines the C2NP benchmark pipeline, including dataset construction and evaluation flow, and Figure 2 visualizes representative crystal-to-nanoparticle transformations derived from DFT-validated unit cells. 2 Related Work 2.1 Bulk Crystals and Finite