This article presents a study on low-index mesoscopic surface reconstructions of gold (Au) surfaces using Bayesian machine-learned force fields. The research addresses the challenge of simulating surface reconstructions, which are difficult to capture with traditional methods like density functional theory (DFT) or classical force fields. By employing active learning of Bayesian machine-learned force fields trained from ab initio calculations, the study enables large-scale molecular dynamics (MD) simulations to describe the thermodynamics and time evolution of low-index surface reconstructions of Au, such as the Au(111)-'Herringbone', Au(110)-(1×2)-'Missing-Row', and Au(100)-'Quasi-Hexagonal' reconstructions. These simulations provide direct atomistic understanding of the dynamic emergence of these surface states, including previously inaccessible information such as nucleation kinetics and a complete mechanistic interpretation of reconstruction under strain and local deviations from stoichiometry. The study successfully reproduces previous experimental observations of reconstructions on pristine surfaces and provides quantitative predictions of spinodal decomposition and localized reconstruction in response to strain at non-ideal stoichiometries. A unified mechanistic explanation is presented of the kinetic and thermodynamic factors driving surface reconstruction. Additionally, the study examines surface reconstructions on Au nanoparticles, where characteristic (111) and (100) reconstructions spontaneously appear on various high-symmetry particle morphologies. The research highlights the importance of accurate surface modeling in materials science, as surface structures significantly influence device performance and stability. The study demonstrates the effectiveness of machine learning force fields (MLFFs) in capturing surface reconstructions, offering a flexible model that can be learned directly from ab initio training data and enables MD simulations at high computational efficiency. The MLFFs are validated against experimental observations and provide insights into the mechanisms and kinetics of surface reconstructions. The study also addresses fundamental questions about the factors influencing surface reconstruction, including the role of initial surface structure, applied strain, adatom concentration, and surface stoichiometry. The results provide benchmarks for future experimental studies and contribute to the understanding of catalyst synthesis, pretreatment, and reactivity control. The study demonstrates the potential of MLFFs in computational investigations of surface reconstruction phenomena and their interpretation in surface science.This article presents a study on low-index mesoscopic surface reconstructions of gold (Au) surfaces using Bayesian machine-learned force fields. The research addresses the challenge of simulating surface reconstructions, which are difficult to capture with traditional methods like density functional theory (DFT) or classical force fields. By employing active learning of Bayesian machine-learned force fields trained from ab initio calculations, the study enables large-scale molecular dynamics (MD) simulations to describe the thermodynamics and time evolution of low-index surface reconstructions of Au, such as the Au(111)-'Herringbone', Au(110)-(1×2)-'Missing-Row', and Au(100)-'Quasi-Hexagonal' reconstructions. These simulations provide direct atomistic understanding of the dynamic emergence of these surface states, including previously inaccessible information such as nucleation kinetics and a complete mechanistic interpretation of reconstruction under strain and local deviations from stoichiometry. The study successfully reproduces previous experimental observations of reconstructions on pristine surfaces and provides quantitative predictions of spinodal decomposition and localized reconstruction in response to strain at non-ideal stoichiometries. A unified mechanistic explanation is presented of the kinetic and thermodynamic factors driving surface reconstruction. Additionally, the study examines surface reconstructions on Au nanoparticles, where characteristic (111) and (100) reconstructions spontaneously appear on various high-symmetry particle morphologies. The research highlights the importance of accurate surface modeling in materials science, as surface structures significantly influence device performance and stability. The study demonstrates the effectiveness of machine learning force fields (MLFFs) in capturing surface reconstructions, offering a flexible model that can be learned directly from ab initio training data and enables MD simulations at high computational efficiency. The MLFFs are validated against experimental observations and provide insights into the mechanisms and kinetics of surface reconstructions. The study also addresses fundamental questions about the factors influencing surface reconstruction, including the role of initial surface structure, applied strain, adatom concentration, and surface stoichiometry. The results provide benchmarks for future experimental studies and contribute to the understanding of catalyst synthesis, pretreatment, and reactivity control. The study demonstrates the potential of MLFFs in computational investigations of surface reconstruction phenomena and their interpretation in surface science.