Rediscovering Armstrong’s quinoid theory with machine learning using quantum chemistry

Abstract

Throughout the history of chemistry, human efforts to design functional molecules have caused the discovery of numerous theories. Recently, artificial intelligence (AI)-enabled de novo molecular generators (DNMGs) have automated molecular design based on data-driven or simulation-based property estimates, eliminating the need for chemical-theory-based guidelines. However, it is unclear whether these DNMGs can discover theories that elucidate molecular design and chemistry. Herein, we demonstrate that an AI-enhanced DNMG can discover a chemical theory regarding the molecular structure. We attempted to elucidate the theory used by the DNMG to generate pure organic molecules (consisting of H, C, N, and O) for absorbing long-wavelength light by observing the functional group enrichment of molecules during the density-functional-theory-based generation. Although the DNMG initially targeted diketone and aniline derivatives, it later focused on quinone derivatives with a long absorption wavelength, as if it found a rule for chemical constitutions relevant to colour, previously known as Armstrong’s quinonoid theory, which claimed that the colour originates from 1,4-quinon derivatives. Additionally, the DNMG shows the potential of 1,2-quinone derivatives as chromophores, as demonstrated by our experimental validation by synthesising one mimetic generated molecule. This study confirms that DNMGs have the potential to discover and expand chemical theories.