Computer- aided formulation development - Computer-Aided Applications in Pharmaceutical Technology - page 26

appropriate selection of network architecture was the milestone in
utilization of ANNs. A Generalized Regression Neural Network (GRNN),
MLP, and Radial Basis Function (RBF) ANN architectures (Statistical
Neural Networks, StatSoft, Inc., Tulsa, OK, USA) were used throughout
the study. In the presence of the Plurol ® Isostearique cosurfactant, a feed-
forward GRNN comprising four layers (the fi rst layer had 2 input units,
the second layer had 27 hidden units, the third layer had 2 units, and the
fourth layer had 1 output unit), was characterized by the generalization
ability of 99.1%. When Cremophor ® RH40 was used as a cosurfactant, a
MLP network with 4 layers was generated with the prediction ability of
92% for training data set, 93% for validation data set, and 92% for test
data set. In systems with Solubilisant gamma ® 2421 and Solubilisant
gamma ® 2429, satisfactory results were achieved with the RBF network.
The ANN models provided a deeper understanding and prediction of a
water solubilization limit for any combination of surfactant concentration
and oil concentration in their mixture, within the investigated range.
Learned networks were used for modeling, simulation, and optimization
of the microemulsion area boundary by testing experimental points in
experimental fi elds; searching for the optimal solutions; and presenting
response surfaces (or contour plots). Response surfaces presenting the
infl uence of the surfactant concentration in the surfactant/cosurfactant
mixture and the oil concentration in the mixture with tensides on the
water solubilization limit, pointed to the maximum performance in the
presence of Cremophor ® RH40 at high SCoS/O ratios (SCoS/O >7:3)
within the investigated K m range. Such mixtures would be the most
promising regarding the self-microemulsifi cation phenomenon. The
combination of the titration method for phase behavior data collection
with in silico data modeling, demonstrated in this study, is a particularly
useful approach in development of SMEDDS, which allows to follow
dilution of self-microemulsifying concentrate with the aqueous phase in a
continuous manner.
The study of Podlogar et al. (2008) demonstrated that ANN modeling
could be effective in minimizing the experimental efforts characterizing
complex structural features of microemulsions. Two evolutionary ANNs
(Yao, 1991) have been constructed by introducing GA to the feed-forward
ANN, one being able to predict the type of microemulsion from its
composition and the second to predict the type of microemulsion from
the differential scanning calorimetry (DSC) curve. The components of the
microemulsion-forming system were isopropyl myristate (oil),
polyoxyethylene (20) sorbitan monopalmitate (Tween ® 40) (surfactant),
glyceryl caprylate (Imwitor ® 308) (cosurfactant), and twice distilled

￿

￿ ￿