(140e) Optimizing Drying Profile of Polymeric Drug Products Using Machine Learning and First Principle Modeling

Palanisamy, A., Dr. Reddy's Laboratories
Mehan, R., Dr Reddys Laboratories
Garikipati, S. J., Dr Reddys Laboratories
Palaparthi, R., Dr Reddys Laboratories

Drying is one of the critical
steps for manufacturing of polymeric drug products. The key quality attributes
for the dried product can be the residual solvent content, and the flowability.
The flowability of the powders can be dependent on the glass transition
temperature which in turn depends on the solvent content. Hence it requires
special precautions to keep the temperature of the drug product during drying below
a critical transition temperature (Tg). This work shows a case study where such
type of issues can be tackled by employing a hybrid model combining machine
learning and first principles to optimize the drying profile in minimal number
of experimental runs.

Typically the drug product is
dried in an agitated Nutsche filter type of a dryer. This drying requires
special care to ensure that at all points of drying; the temperature of the product
is below the critical value (Tg ).  While a simple heat transfer model (like in
Figure 1, section A) can be used to link the jacket temperature to the product
temperature, the presence of solvent and its evaporation needs additional
considerations as detailed below. In the model building step, Model 1 was a
Support Vector Machine (SVM) based model to predict the Tg for different
residual solvent levels (based on experimental data). The model predicted with
accuracy over 98%. Model 2 captures the mass transport of the solvent from the
drug product, accounting for the temperature dependent mass transport of the


Here Ki and Ci
are the diffusivity and concentration of solvent i, and T is the product temperature.
 is calculated as a function
of temperature of each solvent in the system, based on the regression of the
experimental data.

Figure 1: Modeling approach for drying

To predict the optimal jacket temperature
profile for drying to meet the product requirements, the heat transfer model, Model
1 and Model 2 are solved using an algorithm like the one shown in Figure 2. For
a given temperature profile, and initial solvent levels, Model 2 predicted the
time evolution of the solvent concentration. At each time, Model 1 calculated
the Tg. The optimal input temperature profile was chosen such that the product
temperature at any time was lower than the predicted Tg at that time.

Figure 2: Modeling algorithm for
predicting the optimal temperature profile

The above model building approach
of combining machine learning and first principle modeling can be extended to
other unit operations as well where only part of the underlying physics is
understood. In combination with a mixing type of a model, this approach can be applied
to scale-up drying (Figure 1). The presentation shows specific example cases of
such applications.