Abstract:The description and analysis of chemical bonds have been difficult following the popularization of electronic structure calculations. Although many attempts have been made from the perspective of electronic structure, the sheer volume of information in the electronic structure has left contemporary chemical bond analysis methods grappling with an inescapable "Trilemma" where the model briefness, generality, and descriptiveness (descriptive power) cannot be obtained simultaneously. To push the generality and descriptiveness to their extremes, herein a general machine learning-based framework is introduced to compact chemical bonds into a detailed residue-by-residue "genome" with matched encoding/decoding tools. The framework fuses the quantum mechanical aspects, auto feature extraction, nanostructures and/or simulations, and generative models. The encoded genomes are information-dense and decodable, where 100% generality is guaranteed. The descriptiveness of genomes appears to be broader than most known models. As a proof of concept, the realization presented in this work compacts the complete information regarding two critical chemical bonds in thiolate-protected gold nanoclusters, the S-Au and Au-Au bonds, from a Bosonic-Fermionic character perspective into 8-valued genomes. The machine learning component is trained based on 26,528 density functional theory simulated electron localization function images. With an exploration of the space span for the genome, bond polarization, hybridization, intrusion of other atoms, alignments, crystal orientation, atomic motions, and more details are observed. Furthermore, it has emerged from extensive generation tests that molecules and solids can be integrated in such a concise manner than is typically achieved with purely geometric representations. To showcase the intraclass complexity of S-Au and Au-Au bonds visually, a roadmap is plotted by summarizing and correlating the similarities of 8-value-genomes. Furthermore, genomes can be associated with realistic indices easily with a simple multilayer perception architecture as a simple calculating tool. Besides, there are 3 sets of applications, including a set of chemisorption, a set of molecular dynamical analysis, and a set of ultrafast processes, showcasing the interpretability potentials of interatomic genomes in the geometric structures, kinetic properties, and vibration characteristics of molecular systems. As the framework rose to the challenge of nanoclusters from a complicated mesoscopic family of material, the displayed generality and comprehensiveness indicate that the model may "understand" chemical bonds in a machine's way.

Polyatomic Complexes: A topologically-informed learning representation for atomistic systems

Completeness of Atomic Structure Representations

Unified theory of atom-centered representations and message-passing machine-learning schemes

Quantum 3D graph structure learning with applications to molecule computing

From structure mining to unsupervised exploration of atomic octahedral networks

Flexible Topology: A Dynamic Model of a Continuous Chemical Space

Molecular set representation learning

GTAM: A Molecular Pretraining Model with Geometric Triangle Awareness

Unified 2D and 3D Pre-Training of Molecular Representations

Learning Multi-view Molecular Representations with Structured and Unstructured Knowledge

Learning Harmonic Molecular Representations on Riemannian Manifold

Substructure-Atom Cross Attention for Molecular Representation Learning

Topological Characterization of Metal-Organic Frameworks: A Perspective

Molecular representations for machine learning applications in chemistry

Atom-Density Representations for Machine Learning

Active Causal Learning for Decoding Chemical Complexities with Targeted Interventions

Representation of molecular structures with persistent homology for machine learning applications in chemistry

Molecular CT: Unifying Geometry and Representation Learning for Molecules at Different Scales

Generalized Representative Structures for Atomistic Systems

A physics-inspired approach to the understanding of molecular representations and models

Pursuing Extreme Descriptive Power and Generality in Chemical Bond Theories: A Method to Decipher "Interatomic Genomes" from Interatomic Electron Structures