Enhanced ribbon quality in roller compaction process by mitigating splitting through a machine-learning framework.

Area of Science:

• Pharmaceutical Manufacturing
• Process Engineering
• Data Science in Pharmaceuticals

Background:

• Ribbon splitting during roller compaction compromises granule uniformity and tablet quality.
• Predicting ribbon splitting is challenging due to complex process variable interactions.
• Current methods lack the precision to fully understand and control this phenomenon.

Purpose of the Study:

• To develop a machine learning framework for modeling and characterizing ribbon splitting in dry granulation.
• To improve the prediction of ribbon quality and identify optimal process conditions.
• To enhance the interpretability of ribbon splitting phenomena using feature importance analysis.

Main Methods:

• Utilized a Gaussian process regression (GPR)-based neural network with transfer learning.
• Employed SHapley Additive Explanations (SHAP) for model interpretability and feature importance.
• Leveraged multivariate experimental data for model training and validation.

Main Results:

• Achieved reliable predictions of ribbon quality (thickness, density) with high R² and low MSE/MAE.
• Identified feasible operating regions and optimal conditions for consistent product quality.
• Demonstrated the framework's flexibility, scalability, and generalizability across conditions.

Conclusions:

• The developed machine learning framework effectively models and predicts ribbon splitting.
• The approach facilitates early fault detection, accelerates process development, and supports QbD.
• Enables actionable insights for closed-loop control in pharmaceutical dry granulation.