Академический Документы
Профессиональный Документы
Культура Документы
A R T I C L E I N F O A B S T R A C T
Article history:
Received 29 November 2016 Computational modelling of twin-screw granulation was conducted by using an articial neural network
Received in revised form 1 February 2017 (ANN) approach. Various ANN congurations were considered with changing hidden layers, nodes and
Accepted 2 February 2017 activation functions to determine the optimum model for the prediction of the process. The neural
Available online 3 February 2017 networks were trained using experimental data obtained for granulation of pure microcrystalline
cellulose using a 12 mm twin-screw extruder. The experimental data were obtained for various liquid
Keywords: binder (water) to solid ratios, screw speeds, material throughputs, and screw congurations. The
Computational modelling granulate particle size distribution, represented by d-values (d10, d50, d90) were considered the
Model predictive control
response in the experiments and the ANN model. Linear and non-linear activation functions were taken
Wet granulation
into account in the simulations and more accurate results were obtained for non-linear function in terms
ANN
Continuous pharmaceutical manufacturing of prediction. Moreover, 2 hidden layers with 2 nodes per layer and 3-Fold cross-validation method gave
the most accurate simulation. The results revealed that the developed ANN model is capable of predicting
granule size distribution in high-shear twin-screw granulation with a high accuracy in different
conditions, and can be used for implementation of model predictive control in continuous
pharmaceutical manufacturing.
2017 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijpharm.2017.02.009
0378-5173/ 2017 Elsevier B.V. All rights reserved.
S. Shirazian et al. / International Journal of Pharmaceutics 521 (2017) 102109 103
Table 1
Design of experiments for ANN simulation of wet granulation.
Run L/S Screw speed (rpm) Powder ow rate (g/h) Screw conguration d10 (Micron) d50 (Micron) d90 (Micron)
1 0.54 64 49.75 2 kneading zones 15.07 88.535 328.8
2 0.94 200 98 2 kneading zones 151.35 417.2 1004.15
3 1.22 86 98.2 2 kneading zones 357.35 759.9 1202
4 0.54 200 69.33 2 kneading zones 23.885 130.2 360.3
5 1.22 200 61.47 2 kneading zones 307.15 689.95 1038.7
6 1.22 50 82 2 kneading zones 445.65 767 1105
7 1.21 200 49.4 1 kneading zone 385.75 1008.2 1303
8 0.78 200 49.75 1 kneading zone 65.195 271.6 653.5
9 1.21 115 49.5 1 kneading zone 455.15 1052.65 1313.5
10 0.49 185 67.55 1 kneading zone 14.605 67.46 342
11 0.65 50 97.6 1 kneading zone 10.57 76.865 342.6
12 0.52 50 97.05 1 kneading zone 8.33 35.805 300.5
13 1.22 50 82 1 kneading zone 177.5 440.55 831.8
14 0.55 200 98.6 conveying elements only 25.06 113.25 291.35
15 1.22 200 98.15 conveying elements only 332.55 860.65 1250
16 0.76 200 99.2 conveying elements only 66.91 241.2 641.45
17 0.59 85 51 conveying elements only 35.02 164.7 384.45
18 1.12 72 53.25 conveying elements only 186.775 367.975 820.125
19 0.48 50 49.8 conveying elements only 18.42 102.745 341.2
20 0.50 200 95.05 2 kneading zones with cutting elements 7.585 22.425 113.75
21 1.20 58 82 2 kneading zones with cutting elements 330.7 865.6 1273.5
22 0.66 200 49.65 2 kneading zones with cutting elements 35.605 188 490.7
23 0.54 50 61.45 2 kneading zones with cutting elements 14.875 87.55 342.7
24 1.22 200 97.38 2 kneading zones with cutting elements 325.4 763 1154
104 S. Shirazian et al. / International Journal of Pharmaceutics 521 (2017) 102109
applied the computational intelligence-based models for simula- The experiments were designed based on a Custom-designed
tion of oscillating milling. Genetic programming was coupled with DoE considering feed ow rate, screw speed, liquid to solid ratio (L/
ANN in order to predict the particle size during the milling process. S), and screw congurations as factors and particle size distribu-
The model capability in the prediction of d50 values conrmed that tion in terms of d-values as a response. The designed experiments
the ANN is a powerful tool for simulation of pharmaceutical and the obtained d-values of particle size are listed in Table 1. Fig. 1
processes specically for complex unit operations. illustrates the photo of two screws used in this work, and the screw
Therefore, there is a requirement for the development of a conguration can be seen schematically in the Supplementary le.
robust, reliable and fast model to be used for the prediction of high- The obtained granules were dried overnight in an oven at 40 C
shear wet granulation. The main aim this study is to develop a and then characterized using dry powder dispersion laser
methodology to predict the wet granulation process in an efcient diffraction (Microtrac S3500). As the size limit for Microtrac
manner. The predictive method is based on the simulation of this S3500 is 1.4 mm, a sieve was used prior to laser diffraction to
process using articial neural network. An effort is made to build a measure the mass of material greater than 1.4 mm. Consequently
robust, reliable, and fast model of processing which can be used for the recorded d-values of particle size are for the mass fraction
implementation of model predictive control in pharmaceutical smaller than 1.4 mm. As the% mass fraction > 1.4 mm was usually
manufacturing. Both linear and non-linear models are investigated less than 1%, this was considered adequate.
at different network topologies to nd the optimum network for
wet granulation. The model predictions are then compared with 3. Modelling
experimental data obtained on a 12 mm twin-screw extruder. To
the best of our knowledge, there is no work on ANN modelling of 3.1. ANN conguration
high shear twin-screw granulation of pharmaceutical formula-
tions. For the development of the ANN model, in addition to screw
conguration, three process parameters including liquid to solid
2. Experiments ratio (L/S), screw speed, and powder ow rate were considered as
inputs. On the other hand, d-values of particle size distribution
Wet granulation experiments were conducted via a twin-screw including d10, d50, and d90 were considered as response
extruder, ZE12 (ThreeTec, Switzerland) with screw diameter of parameters. Basically, ANN models consist of three layers; i.e.
12 mm and length-to-diameter ratio 40:1. Microcrystalline cellu- input, hidden and output layers (Kazemi et al., 2016). Important
lose (MCC, Avicel pH 101 Pharmatrans) was used as a formulation parameters in development of ANN include: validation method,
model and water as a binder throughout the experiments. A activation function, hidden layers, and number of nodes in each
peristaltic pump (Watson Marlow 520SN) was used for adding layer of network. These parameters were taken into account to nd
water and a gravimetric feeder (Three-Tec, Switzerland) for feeding out the optimum ANN topology for predicting the process. Finding
the powder. the optimum ANN topology is challenging since there is no
straightforward procedure for this, and usually a trial and error
Table 2
Comparisons between different ANN topologies in terms of nodes and activation
functions.
approach is utilized for building an appropriate model. Kfold show the model accuracy. Coefcient of determination which is
method was used as validation in which K = 3 showed the best dened as (Barrasso et al., 2015):
results. The latter means that the whole experimental measure- X 2
ments are divided into three sections, and each set of data is 2
f i yi
R 1 Xi 1
evaluated for validation and the best t is selected. Therefore, y yi 2
i i
among 24 experimental runs listed in Table 1, 16 runs were used for
training of the network and 8 runs were used to validate the built where R2 refers to coefcient of determination. f and y refer to
ANN model. JMP Pro 12 software was used for ANN modelling and predicted and observed values, respectively. i also refers to the set
analysis of the results. The software is capable of considering of experimental runs.
categorical factors as input for ANN tting. Moreover, the goodness of ANN t is reported as root-mean-
squared error which is dened as follows:
3.2. Model assessment sX
2
f yi
i i
RMSE 2
The developed ANN model is used for prediction of d-values in n
twin-screw granulation in which the predictability of model is where n refers to the number of measurements
assessed through comparing the predicted values with observed
(experimental) values. Two statistical parameters were used to
e2x 1
4.1. Experimental results 3
e2x 1
The experimental data obtained from measuring d-values of where x is a linear combination of input variables
particle size distribution are reported in Table 1. As seen, 24 Up to four nodes were considered for each layer (two linear and
experiments were run and d-values for each run were measured two non-linear) and the results were compared in terms of average
and reported as d10, d50, and d90. Rather than tting the ANN RMSE for overall training and predictions. Increasing the number of
model to all d-values and particle sizes, these three d-values were nodes greater than two increased the solution time and no
chosen to t the model to avoid the complexity of objective signicant improvement was observed. Since the objective
function in parameter estimation. Moreover, for industrial function may have different local minima, multi-start technique
applications the granule size range are of great importance rather was used for parameter estimation and tting. This means that the
than the whole particle size distribution (Barrasso et al., 2015). ANN tries to estimate the tting parameters by different initial
points in order to nd the global minimum of objective function.
4.2. Determination of hidden layers and nodes The iteration parameter was set to be 10 for multi-start technique
(Kazemi et al., 2016).
Basically, ANN consists of layers and nodes in which increasing The results are listed in Table 2. As shown, the ANN topology
hidden layers and notes would make the ANN model more exible. with two nodes for both hidden layers considering a combination
In this work, two hidden layers and up to four nodes per layer were of non-linear and linear activation functions shows the best
considered for the development of ANN. However, more nodes and performance in tting the experimental data. However, the results
layers could be considered but the solution time would increase are not still satisfactory given that the model could be used as
considerably. Since the main aim of ANN model in this work is to be reliable tool for prediction of process.
used for MPC application, the solution time is the most important
criterion for ANN development. For each hidden layer, an 4.3. Boosting the network
activation function is required to estimate the value of response.
The activation function applies at the nodes of hidden layers and is The ANN developed so far with two hidden layers and two
a transformation of a linear combination of input variables. For this nodes per layer has shown to be weak for accurate prediction of
work, two different activation functions are considered including granule size in twin-screw granulation. Considering this, an
linear and non-linear. For linear activation function, the linear attempt was made to build a more robust and reliable model to
combination of input variables are not transformed whereas for predict the process. To do this, the trained networks were made
non-linear, the hyperbolic tangent function is used as follows more robust by boosting approach. Boosting is used of construct
Model R2 Ref.
4.5. Effect of process parameters on granule size
Multi-dimensional PBM 0.805 Barrasso et al. (2015)
Coupled ANN-PBM-DEM 0.97 Barrasso et al. (2014)
Once the ANN model is validated, it can be used as a predictive ANN 0.99 (This work)
tool to evaluate the effect of various process parameters and screw
108 S. Shirazian et al. / International Journal of Pharmaceutics 521 (2017) 102109
Fig. 7. Effect of process parameters and screw congurations on span of particle size distribution.
S. Shirazian et al. / International Journal of Pharmaceutics 521 (2017) 102109 109