Explainable artificial neural network as a soft sensor to predict the moisture content in a continuous granulation...

Area of Science:

• Pharmaceutical Manufacturing
• Chemical Engineering
• Data Science

Background:

• Artificial neural networks (ANNs) offer potential for pharmaceutical manufacturing but face trust issues due to their black-box nature.
• Soft sensors are crucial for real-time process monitoring, but traditional methods can be limited.

Purpose of the Study:

• To develop and evaluate an explainable artificial neural network (ANN) as a soft sensor for monitoring continuous manufacturing processes.
• To estimate the moisture content of granules using only process parameters, without direct measurement.
• To compare ANN performance with traditional near-infrared (NIR) spectroscopy methods.

Main Methods:

• Development of two ANNs: a multilayer perceptron (MLP) and a Nonlinear Autoregressive with Exogenous Inputs (NARX) model.
• Utilizing process parameters as inputs to predict moisture content.
• Employing SHAP analysis to investigate the explainability of the MLP model.
• Validation using off-line loss-on-drying measurements and comparison with NIR spectroscopy.

Main Results:

• ANN models (MLP and NARX) achieved comparable accuracy to NIR spectroscopy for moisture content prediction.
• Moisture content was determined with a root mean square error of prediction below 1% across all methods.
• SHAP analysis provided transparency into the MLP model, identifying key parameters influencing predictions.

Conclusions:

• Explainable ANNs are a viable and cost-efficient alternative to traditional analytical methods for soft sensing in pharmaceutical manufacturing.
• The developed ANN soft sensors enhance process understanding and reliability in continuous manufacturing.
• This approach offers an orthogonal and transparent method for monitoring critical quality attributes.