Models building models for discovery of graphene metamaterial design principles in the context of failure
Year: 2026
Type: preprint
Abstract: Artificial intelligence assists scientific discovery, but most agents operate previously developed instruments. Whether an agent can build an instrument, translating physical theory into reliable measurements and testable predictions, then use it for open-ended research remains open. Here, from a single prompt, an AI agent built and validated an atomistic simulator and conducted 96 simulations across ten graphene architecture families. At matched mass, load paths governed strength, and tilted slits revealed distinct failure mechanisms. Human–AI collaboration examined larger strains and hierarchical architectures. Sliding slit edges exhibited sequential geometric reconnection; hierarchical veins redistributed load, compartmentalized damage and deflected cracks. Selected larger hierarchical designs were about 25% stronger, with at least one-third greater work to failure, than equal-mass, equal-size controls. These results demonstrate a workflow where AI agents build scientific models and use reflective reasoning to falsify, validate and refine physical understanding, while human–AI collaboration drives mechanistic discovery and engineering of complex materials.
Source: ChemRxiv
Author: Markus J. Buehler
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7215077882