Tracing the Physical Roots of Chemical Concepts through Artificial Intelligence
- 1Dipartimento di Scienze Chimiche, Università degli Studi di Padova, Via Marzolo 1, Padova, Italy, 35131
- 2Interdisciplinary Center for Scientific Computing, Ruprecht-Karls University, Im Neuenheimer Feld 205, Heidelberg, Germany, 69120
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.
Related
- Chemical bonding concepts emerge naturally from maximally entangled atomic orbitals
- Chemical Feedback and Collective pH-Dynamics in Enzyme-Powered Artificial Cells.
- The Role of Electron-Nuclear Coupling on Chemical and Physical Processes
- Using artificial intelligence for clinical trial quality assurance in a multi-institutional neurovascular-sparing prostate radiotherapy trial
- Future of Artificial Intelligence for Science in Japan 2024 Community Report
BibTeX
@misc{privat2026tracing,
title = {{Tracing the Physical Roots of Chemical Concepts through Artificial Intelligence}},
author = {Erwan Privat and Giovanni Parolin and Ajay M. S. Rawat and Antonino Polimeno and Sergio Rampino},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010384/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010384/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Tracing the Physical Roots of Chemical Concepts through Artificial Intelligence
AU - Erwan Privat
AU - Giovanni Parolin
AU - Ajay M. S. Rawat
AU - Antonino Polimeno
AU - Sergio Rampino
PY - 2026
DA - 2026/10/13
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010384/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010384/v1
AB - 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.
ER -