(603d) Applying Acoustic Emissions for Real-Time Monitoring of Particle Size in Twin Screw Granulation with Inelastic Formulations

PURPOSE: Twin screw granulation (TSG) generates a bimodal particle size distribution (PSD) that is difficult to monitor due to a lack of process analytical technology (PAT). An inline technique based on ultrasonic acoustic emissions (AE) was originally proposed for this purpose and tested with a lactose formulation. A unique filter was created to overcome interference arising from auditory masking, with the resulting predicted PSD showing less than 1% error [1]. However, most pharmaceutical formulations are much more inelastic than pure lactose for their collisions generating AE and in these cases, the earlier filter was inadequately correcting for masking. This presentation covers the development of a new filter that accounts for inelastic forces through the review and selection of an appropriate micromechanical models and thus improve the accuracy of predicted PSD for a broader range of formulations.

PSD PREDICTIVE MODEL:

The proposed PAT system is based on an artificial neural network (ANN) architecture containing an input layer, four hidden layers, and one output layer. The Rectified Linear Unit activation function (ReLU) was used for the first and fourth hidden layer, whereas the second and third hidden layers had the hyperbolic tangent (tanh) and sigmoid function, respectively. Finally, the output layer had a linear activation function. To quantify prediction error in the model, the root mean squared error (RMSE) was used. This architecture and choice to use AI was made because its ability to handle highly non-linear systems when compared to statistical modelling.

Over the course of the study, three separate ANN were trained with the above architecture. The first two ANN were trained only with the AE data and PSD of granulated lactose monohydrate trials as the elastic case, transformed with either the spring and dashpot (SPD) or Walton-Braun (WBR) filter (to be described below). The predicted PSD based on AE data of three other formulations (K4M, F4M, and KSR) of varying inelasticity were then studied with these two elastic ANN. The purpose of this stage in the study was to examine how the adjustable parameter of each filter improved the predictions relative to the elastic case. The average RMSE from each of those predictions was plotted as a function of the adjustable parameter to showcase how varying viscous or plastic contributions corrected to interpreted AE signal from the impact event based on the degree of inelasticity exhibited by the granules; inelasticity was varied in the experiments based on formulation and moisture content. This analysis ultimately determined the preferred transformation to be used for the modified PAT. After the preferred transformation was determined, the third ANN was trained with data based on all formulations to demonstrate the predictive accuracy of the new PAT.

METHODS: Four formulations were granulated in a 27 mm Leistriz ZSE HP twin-screw granulator at various temperatures and liquid-to-solids (L/S) ratio with the experimental conditions shown in Table 1 (attached). The excipients used in this study were Flowlac® 100 (Lactose monohydrate), Avicel PH102, METHOCEL^TM K4M, METHOCEL^TM F4M and Kollidon SR®. Exiting wet granules fell 20 cm to strike a stainless steel plate coupled to a broadband AE sensor. Granulated samples were air-dried for 48 hours to 5% moisture content and then classified using a sieve shaker to generate a 16-point particle size distribution ranging from 44 microns to 7.15 millimeters. The AE and PSD data was split into 80% training and 20% testing, and 10% of the training data was used for validation during model training. All ANNs were setup in Python using the keras library.

A high speed camera (recorded at 40,000 fps) was used to determine the coefficient of restitution (COR) of each formulation for quantifying their inelasticity of impact. Results showed K4M was the most inelastic formulation (coefficient of restitution, COR=0.24), followed by F4M, KSR, and Lac (COR=0.79).

The pre-processing of the AE signal involved a discreet Haar wavelet filter followed by using a fast Fourier transform to analyze the signal in the frequency domain. Two micromechanical models were utilized to describe the impact response of the granules impacting the plate, including:

1) Non-linear Hertzian spring and dashpot model (SPD) which assumes a viscoelastic response

2) Walton-Braun model (WBR) which assumes an elastoplastic response

A descriptor of each micromechanical model was taken as an adjustable parameter to be used in the modified filter to be able to account for the different inelastic forces in each formulation. A dampening coefficient and the COR were the parameters of interest for the SPD and WBR models, respectively. The purpose of choosing either micromechanical model in this case was to describe improvements in predictive accuracy through physical attributes such as the degree of inelasticity associated with water’s influence on the different components of the granule, or the production of viscous liquid bridges of the granules.

The models were then introduced to the original auditory masking filter, creating two separate transformations before being used for model training/validation through artificial neural networks (ANN). The dampening coefficient was varied from 1 x 10^-6 to 1000 to account for non-dampened to fully dampened conditions, whereas the COR was varied from 1 x 10^-6 to 1 to showcase elastic to fully inelastic behaviour.

RESULTS

The viscoelastic SPD filter with the elastic ANN model ultimately showed minimal effect on K4M, a decrease in error with KSR and an increase in error with F4M with increasing dampening parameter. The differing trends with the various formulations showed the viscoelastic filter could not robustly account for how the AE signal was affected by the nature of the different granules’ collisions.

The elastoplastic WBR filter with the elastic ANN showed a more consistent response to all formulations tested, with the highest RMSE being at the lowest tested COR of 1×10^-6 in all cases. The consistent trend and improvement in error made by adjusting COR for each of the three inelastic formulations showed the collisions of granules with the impact plate were better described of elastoplastic and that the WBR model was more robust in improving the fit of varying inelastic formulations to the elastic ANN’s understanding of AE events based on particle size. Therefore, the Walton-Braun model was chosen as the preferred transformation for the improved PAT and training of the inelastic ANN.

After training the inelastic ANN with AE and PSD data from all formulations, the improved PAT system was compared the original (elastic) PAT using test specimens (never used in training) from the entire set of formulations. The predictive accuracy showed large improvements after applying the elastoplastic WBR filter, going from a maximum RMSE of 18.6 wt% for the elastic PAT to a maximum of 2 wt% for all particle sizes measured. This outcome matches previous findings while studying lactose granulation, that training the ANN model had the strongest effect on improving predictive accuracy but only with the correct filter used was it possible to reduce the error for all particle sizes.

CONCLUSIONS:

Multiple typical pharmaceutical formulations prepared by wet granulation with a twin-screw granulator were explored to develop an improved PAT system based now on the principles of inelasticity to more accurately predict the PSD during inline monitoring. Comparing the spring and dashpot as well as the Walton and Braun micromechanical collision models for an auditory masking filter, it was found that AE generated by these wet granules was most robustly described by their elastoplastic behaviour. Training a new ANN based on this broad set of formulations from the elastic (lactose) to highly inelastic (K4M formulation) produced a robust predictive system for samples exhibiting bimodal particle size distribution. Future work for developing this PAT will focus on adapting the filter to include aspect ratio (as most particles produced from TSG are non-spherical in nature) along with expanding the PAT to different processes such as hot melt or dry granulation.

REFERENCES:

[1] H. A. Abdulhussain and M. R. Thompson, “Predicting the particle size distribution in twin screw granulation through acoustic emissions,” Powder Technol., vol. 394.