Roller compaction: Measuring ribbon porosity by terahertz spectroscopy and machine learning

Area of Science:

• Pharmaceutical Manufacturing
• Spectroscopy
• Machine Learning

Background:

• Roller compaction is vital in pharmaceutical manufacturing, with ribbon porosity a critical quality attribute.
• Traditional methods for porosity analysis are slow and provide only average bulk values.
• Terahertz (THz) spectroscopy offers fast, non-destructive porosity measurement but is sensitive to surface topography.

Purpose of the Study:

• To develop and train machine learning models for classifying ribbon topography using THz spectra.
• To improve the accuracy and precision of THz-based porosity measurements in pharmaceutical manufacturing.
• To enable the resolution of density distribution within roller-compacted ribbons.

Main Methods:

• Convolutional neural network (CNN) models were developed and trained.
• THz spectra were used as input for the CNN models.
• Models were trained to classify four ribbon topographies: ridge, valley, flat plane, and edge points.

Main Results:

• Classifiers achieved 91% validation accuracy in identifying outliers and differentiating smooth from knurled surfaces.
• An 81% testing accuracy was achieved for distinguishing between ridges and valleys on knurled surfaces.
• The developed method allows for resolving density distribution within samples.

Conclusions:

• CNN models effectively classify ribbon topography from THz spectra, enhancing porosity analysis.
• This approach addresses the limitations of traditional methods by providing spatially resolved porosity data.
• Combining topography and spectral data allows for accurate average bulk porosity determination compatible with conventional methods.