Air pollution spans metre-scale near-road hotspots to regional and intercontinental transport, yet no single model can represent the full range of processes that control exposure to different pollutants. This review synthesises recent developments in regional-to-local coupling, covering scale-aware regional chemical transport models (CTMs), local dispersion and street-network models and computational fluid dynamics (CFD, from practical RANS to chemistry-enabled LES and emerging GPU/LBM acceleration), and machine learning (ML), including mass-consistent super-resolution, for downscaling and surrogates. One-way “offline” coupling remains the most widely used approach because it is modular and computationally efficient, but its performance depends on how CTM background fields are defined and mapped and on how emission overlap, temporal mismatch and reduced chemistry are handled. Two-way “online” approaches, including Plume-in-Grid (PinG) for point sources and Street-in-Grid (SinG) for dense urban networks, exchange mass during integration, allowing urban plumes, NO–O 3 titration and aerosol–radiation interactions to feedback on regional oxidant budgets and meteorology. These benefits require conservative remapping, turbulence and mixing consistency at the canopy–boundary layer interface, and transparent mapping between chemical mechanisms across scales. Persistent challenges include interface turbulence–chemistry interactions, harmonisation of emissions and meteorological inputs, treatment of urban green infrastructure without double counting drag or deposition, computational feasibility for chemistry/aerosol-coupled CFD, ML transferability under regime shifts, and propagation of input uncertainty. Priority directions include regime-based criteria for when two-way coupling is required, routine mass-budget diagnostics, adaptive or variable-resolution strategies, and ML downscales and surrogates that enforce non-negativity and mass consistency for scenario testing. • Regional-to-local air quality model coupling approaches are investigated across multiple scales. • One-way offline coupling is efficient but sensitive to background and emission overlap. • Online PinG/SinG coupling capture urban-regional feedback but require conservative exchange. • Machine learning downscaling could speed ensembles with mass- and physics-aware constraints.
Dai et al. (Sun,) studied this question.