In chemical reactor engineering, simple concepts are used for describing the flow in a reactor. Turbulent two-phase flow processes, however, are characterized by a broad range of time and length scales. Two-phase reactors operated in the turbulent regime therefore qualify for multiscale modeling. Multiscale models are being developed in such diverging fields as chemical reaction engineering (among which are packed bed reactors, fluidized beds, and risers), chemical vapor deposition reactor modeling, turbulent single-phase and two-phase flow simulations (among which is combustion), and materials science. The common aspects of these multiscale models are highlighted: they all comprise a coarse-grained simulation for the macroscale and some type of fully resolved microscale simulation. Different simulation techniques are used however. Processes involving single-phase flow may require one of three computational fluid dynamics (CFD) techniques: direct numerical simulations (DNSs), large eddy simulations (LESs), and Reynolds averaged Navier−Stokes (RANS)-based simulations. For two-phase flows, two CFD options are open: Euler−Lagrangian (or particle tracking) and Euler−Euler (or two-fluid). The characteristics of all these approaches are discussed. One of the more interesting options in dealing with turbulent two-phase, i.e., multiscale, flow reactors is to run a DNS for the local small-scale processes. Such a DNS is carried out in a periodic box, a dedicated forcing technique being used to impose the turbulent-flow conditions pertinent to a specific position in the macro domain. Several such successful DNSs are reviewed, which all exploit lattice Boltzmann (LB) techniques. Some new and promising results from LB simulations for gas−liquid flow systems are presented. Finally, a truly multiscale simulation strategy is presented for turbulent two-phase flow reactors, which combines a coarse-grained simulation for the macro domain concurrently run with several DNSs for properly chosen positions in the domain. LB techniques are recommended. The crucial step of this strategy—feeding the results of the local DNS frequently back into the coarse-grained simulations until convergence is reached—is described but has not been implemented yet.