Pursuing Extreme\nDescriptive Power and Generality\nin Chemical Bond Theories: A Method to Decipher “Interatomic\nGenomes” from Interatomic Electron Structures

The description and analysis of chemical bonds have been\ndifficult\nfollowing the popularization of electronic structure calculations.\nAlthough many attempts have been made from the perspective of electronic\nstructure, the sheer volume of information in the electronic structure\nhas left contemporary chemical bond analysis methods grappling with\nan inescapable “Trilemma” where the model briefness,\ngenerality, and descriptiveness (descriptive power) cannot be obtained\nsimultaneously. To push the generality and descriptiveness to their\nextremes, herein a general machine learning-based framework is introduced\nto compact chemical bonds into a detailed residue-by-residue “genome”\nwith matched encoding/decoding tools. The framework fuses the quantum\nmechanical aspects, auto feature extraction, nanostructures and/or\nsimulations, and generative models. The encoded genomes are information-dense\nand decodable, where 100% generality is guaranteed. The descriptiveness\nof genomes appears to be broader than most known models. As a proof\nof concept, the realization presented in this work compacts the complete\ninformation regarding two critical chemical bonds in thiolate-protected\ngold nanoclusters, the S–Au and Au–Au bonds, from a\nBosonic-Fermionic character perspective into 8-valued genomes. The\nmachine learning component is trained based on 26,528 density functional\ntheory simulated electron localization function images. With an exploration\nof the space span for the genome, bond polarization, hybridization,\nintrusion of other atoms, alignments, crystal orientation, atomic\nmotions, and more details are observed. Furthermore, it has emerged\nfrom extensive generation tests that molecules and solids can be integrated\nin such a concise manner than is typically achieved with purely geometric\nrepresentations. To showcase the intraclass complexity of S–Au\nand Au–Au bonds visually, a roadmap is plotted by summarizing\nand correlating the similarities of 8-value-genomes. Furthermore,\ngenomes can be associated with realistic indices easily with a simple\nmultilayer perception architecture as a simple calculating tool. Besides,\nthere are 3 sets of applications, including a set of chemisorption,\na set of molecular dynamical analysis, and a set of ultrafast processes,\nshowcasing the interpretability potentials of interatomic genomes\nin the geometric structures, kinetic properties, and vibration characteristics\nof molecular systems. As the framework rose to the challenge of nanoclusters\nfrom a complicated mesoscopic family of material, the displayed generality\nand comprehensiveness indicate that the model may “understand”\nchemical bonds in a machine’s way.