Linzhe Zhang, Changming Xu
The paper introduces a method to automatically discover the fundamental structure of physical measurements—such as intensive/extensive properties, coordinate systems, and dual relationships—from controlled experiments without prior knowledge of the underlying language. It provides a theoretically grounded, polynomial-time algorithm that recovers ontological elements like carrier geometry, measurement types, and residual symmetries, even under noisy, corrupted, or incomplete observations. The approach enables the recovery of complete physical representations, including Maxwell-like structures, from anonymous data.
This work enables machines to autonomously uncover the language of physics from raw experimental data, removing the need for human-defined measurement frameworks. It advances scientific machine learning by shifting focus from learning laws to discovering the representational foundations upon which laws are built. This could accelerate discovery in complex physical systems where human intuition fails.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.