(259c) Expedite Particle Engineering Process Development Using Data Science | AIChE

(259c) Expedite Particle Engineering Process Development Using Data Science

Authors 

Matos, I., Hovione
Vicente, J., Hovione
Semião, V., Instituto Superior Tecnico, University of Lisbon

Expedite
particle engineering process development using data science

T.
Porfirio1,2*, Inês Matos1, J. Vicente1, V. Semião2

8.0pt;line-height:150%;font-family:" arial font-weight:normal>1 line-height:150%;font-family:" arial normal> Hovione Farmaciência SA, Estrada do Paço do Lumiar, 1649-038 Lisbon,
Portugal; *tporfirio@hovione.com

8.0pt;line-height:150%;font-family:" arial font-weight:normal>2 line-height:150%;font-family:" arial normal>LAETA, IDMEC, Mechanical Engineering Department, Instituto Superior
Técnico, Universidade de Lisboa, Av. line-height:150%;font-family:" arial normal>Rovisco Pais 1, 1049-001 Lisbon, Portugal

The scale-up of particle engineering processes such jet
milling, spray congealing or spray drying, is often considered a resource
intensive challenge requiring large quantities of expensive drugs. The
possibility of reducing the experimental burden at large-scale of such
development by using well-structured data set from the prior knowledge is a
very appealing proposition since it not only reduces the development costs but
also the time and material resources. The proposed work uses a hybrid model to link
the operating conditions and process descriptors to the product quality
attributes such particle size and bulk density.

Firstly, the operating conditions are used as input to a
first principles model comprising the key phenomena of the process – thermodynamic,
atomization and congealing kinetics in the case of spray congealing. Process
descriptors are calculated using these models. Further, both process
descriptors and operating conditions are used as input of a black box model.
Different data science algorithms were evaluated as black box: neural networks,
random forest, K-nearest neighbors, support vector machines, and generalized
linear models.

Figure 1
– Observed vs predict of a hybrid model a product quality attribute

Comprehensive case studies of application in jet milling,
spray congealing or spray drying scale-up are explored. Figure 1 shows an
example of the output of the hybrid model. The results of the data science algorithms
are compared with PLS approach. Both data science algorithms and PLS can
successfully predict the particle size from the selected input variables.

Using this approach, the development risk could be reduced
by using a structured tool where the prior experience relates to suitable
models for the prediction of the right descriptors. From this study was
concluded that despite the many influences of process and formulation
variables, trends and correlations could be founded between scale-independent
parameters and the product quality attributes. The established models can be
used to facilitate and expedite the scale-up development for new products in
which lab-scale correlations could increase the degree of prediction.