Although such a description requires details of particle-particle interactions as inputs, which are not easily obtained in experiments, DNS is generally believed to be able to predict the hydrodynamics of well-defined systems reasonably. However, the computational cost involved in DNS is tremendous. In fact, to our knowledge, the largest DNS of gas-solid suspension to date involves only thousands of particles (Ma et al., 2006), and for computationally less demanding liquid-solid systems, the number is still limited to tens of thousands (Nguyen and Ladd, 2005), while for industrial units with dimensions in meters or even ten meters, the number of particles are in the range of 10
and the number of numerical elements needed would reach 10
. In terms of temporal evolution, the time step of the DNS simulations are typically finer than microseconds and the computational time for one step can hardly reach milliseconds even for well parallelized code on state-of-the-art high performance computing (HPC) systems, therefore, an reasonable estimate is “days (computational) for seconds (physical)”. Further development of HPC technology is likely to expand the size of the simulated systems dramatically, but reducing the ratio of computational to physical time will be much more difficult because of the strong barriers in the frequency of chips.
In short, we may say that conventional DNS of industrial CFB units is yet impossible or at least not economical in the foreseeable future, unless certain breakthrough specially designed for parallelization can be made in both software and hardware.
2.2 How accurate is coarse grid simulations?
For reasonable computational cost and timely response, most practical simulations on industrial units have used two-fluid model (TFM) with coarse grids on the scale of several
millimeters to centimeters
. However, as reported by many researchers (Andrews IV et al., 2005; Jiradilok et al., 2006; Lu et al., 2005; Qi et al., 2007; Qi et al., 2000; Sundaresan, 2000; Yang et al., 2003a, 2003b), the accuracy of such simulations is not satisfactory. For example, the solids circulating rates are significantly over-predicted and the meso-scale heterogeneity is considerably smoothed, as discussed in the literature (Yang et al., 2003a). The accuracy of a numerical simulation depends on the rationality of the physical model, the numerical scheme and the grid size used. Assuming that a TFM is reasonable on the scale its averaging procedures are taken, its best solutions can be approached when the grid size is much smaller than this scale, since a smooth and linear change can usually be secured then. When coarser grids are used, the numerical solution will gradually deviate from the analytical solution of the model and the physical picture, at rates largely dependent on the nature of the model and the system simulated. When significant sub-grid scale (SGS) heterogeneity displays in the system, the linear and smooth variation of the flow variable assumed in the numerical solution loses completely, so that the discrepancy between coarse-grid simulations and the picture really described by TFM could be very large, as reported by Sundaresan (2000) and shown in Fig. 2. The calculation of drag coefficient in TFM can serve as a good example where correlations from seemingly uniform suspensions or fixed beds (Wen & Yu, 1966; Ergun, 1952) are routinely used. As previously estimated (Ge, 1998; Li, 2000; Yang et al., 2003b), these correlations may give predictions even with wrong order of magnitude if typical cluster-broth two-phase structures present under the grid scale. This effect can be demonstrated more precisely by Fig. 3 for a suspension of mono-sized 2D particles under different configurations, where the slip velocity induced by a given pressure drop with constant solids concentration is obtained from DNS using macro-scale particle methods (Ma et al., 2006; 2007) with Lagrangian descriptions for both particles and fluid. The naturally developed heterogeneous suspension, though not as significant as in real gas-solid systems owing to the limitation of the number of particles, has quite different slip velocities from that for fixed hexagonal (most uniform) suspension and the ratio of the highest to the lowest drag coefficient is about 6. The simulation also suggests that the form of the heterogeneity, for example, the orientation of the voidage gradient relative to the fluid flow, have great influence on the resistance characteristics of the suspension. Therefore, mean voidage and slip velocity alone are insufficient to define