Year: 2026
Type: preprint
Abstract: The language of chemistry is universally based on simple concepts such as the atom and the bond. Yet, these concepts find no place in a physical description of the nanoscopic world, where–within a Born-Oppenheimer view–a molecule is ultimately represented by a three-dimensional mathematical function expressing the electron density surrounding a given set of nuclei. While several methods have been proposed over the past decades to define atoms and bonds through the analysis of the molecular electron density, consensus on a theory bridging chemical concepts to the underlying physics has not been reached, and this has corroborated the idea that chemical concepts have an intrinsic degree of vagueness. In this work, we harness the power of Artificial Intelligence to trace the physical roots of chemical concepts–from the simple idea of an atom or a bond, to more elaborate ones such as the particular ‘chemistry’ of an atom in a given molecular context–and assess to what extent these can be framed on a quantitative level. Using simple machinelearning models, we show that a computer can be successfully trained to discern the nuanced chemical character of atoms and bonds from compact representations of the electron density with an accuracy of 98 % and 99 %, respectively. These results indicate that elusive and articulated chemical concepts are actually hard-coded in the underlying physics and suggest that chemistry is less ‘fuzzy’ a science than is generally assumed.
Source: ChemRxiv
Institutions: University of Padua, Heidelberg University
Language: English
Subfield: Atomic and Molecular Physics, and Optics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
Funder: Università degli Studi di Padova
OpenAlex ID: W7221671642