Figure 3. Random Forest out-of-bag
permuted predictor importance estimates for local cluster
characteristics. The left column corresponds to 1188 datasets from a
fast fluidized bed 16,17,25, while the right column
corresponds to 378 datasets from a turbulent fluidized bed20.
In an earlier study 26, the 1188 cluster datasets in
the fast fluidized bed were found to be classifiable into two distinct
data assemblies, with one composed of monodisperse particle systems
(i.e., three narrow PSDs) and the other of non-monodisperse particle
systems (i.e., two binary mixtures and one broad PSD), as shown in the
cluster frequency versus duration plot in Figure 4. Clearly, the
clusters from turbulent fluidization appear to be markedly different
from that of fast fluidization, particularly in terms of much higher
cluster frequency. Although the turbulent datasets consist of both
monodisperse and non-monodisperse particle systems, the data points only
overlap with the monodisperse data points of the fast fluidized bed. A
closer look was taken at the data collected from turbulent fluidization
to investigate if the data can be classified into distinct data
assemblies based on the monodispersity versus non-monodispersity of the
particle system, per that observed for fast fluidization (Figure 4).
Figure 5 shows the plots of cluster frequency versus duration, with the
data for the different particle systems represented by different
symbols. Using different colors to represent all nine different particle
systems, Figure 5a indicates no obvious demarcation into distinct data
assemblies. Using three colors to represent the three particle systems
(namely, narrow PSDs, binary mixtures, and broad PSDs), Figure 5b shows
that, relative to the narrow PSDs, the maximum frequency of the binary
mixtures was approximately half of that of the narrow PSDs or broad
PSDs. Figure 5c presents data for the binary mixtures and the two
constituents of narrow PSDs. Although the narrow PSD of polystyrene (PS)
particles gave high-frequency clusters, the binary mixtures did not
exhibit these, but instead gave clusters with larger durations than that
of either constituent. As for the broad PSD mixtures, Figure 5d shows
that the cluster frequencies were largely similar to that exhibited by
the narrow PSD, and a few clusters had durations greater than that of
the narrow PSD. Therefore, whereas the presence of particle mixtures
(either binary mixtures or broader PSDs) gave significantly lower
cluster durations in the fast fluidized bed (Figure 4), the effect is
less marked in the turbulent fluidized bed (Figure 5). This again
affirms that the underlying mechanisms governing cluster formation are
different in the different regimes.