Complexities in the flow of the silicon melt in a Czochralski and a do
uble-layered Czochralski crystal growth system are characterized in te
rms of nonlinear predictions of the melt temperature fluctuations that
are observed at a sampling time of 1 s using thermocouples at various
locations in the melt. The Sugihara-May method is employed as the for
ecasting technique. Information entropy of the thermal sequences is es
timated from the dependence of the normalized root-mean-squared error
of prediction on the prediction-time interval. The dynamical behaviour
of the flows is related to temporally correlated random motion added
to regular motion with scaling exponents specific to the crucible rota
tion rate. The spatial and temporal structures of the flows are, howev
er, totally different between the growth systems. The influence of the
flows on the crystal-growth process is discussed in relation to rando
mness in the thermal sequences.