{"id":"arxiv/2601.19076","url":"https://arxiv.org/abs/2601.19076","tags":["cond-mat.mtrl-sci","cond-mat.mes-hall","cs.LG","quant-ph"],"score":0,"title":"C2NP: A Benchmark for Learning Scale-Dependent Geometric Invariances in 3D Materials Generation","domain":"cond-mat.mtrl-sci","source":"arxiv","authors":["Can Polat","Erchin Serpedin","Mustafa Kurban","Hasan Kurban"],"reviews":0,"subject":"cond-mat.mtrl-sci","summary":"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…","category":"research","comments":0,"disputed":false,"division":"physics","mentions":0,"discussed":0,"published":"2026-01-27","reactions":{}}