Using AI/ML to predict blending performance and process sensitivity for Continuous Direct Compression (CDC).

Area of Science:

• Powder Technology
• Chemical Engineering
• Computational Science

Background:

• Accurate prediction of fill levels in inclined linear blenders is crucial for process control.
• Fill level directly impacts blade passes (strain), a key factor for achieving content uniformity in powder mixtures.
• Existing methods may lack the precision needed for complex powder formulations.

Purpose of the Study:

• To develop and evaluate artificial intelligence (AI) and machine learning (ML) models for predicting fill level in inclined linear blenders.
• To identify the most influential powder characteristics and processing parameters affecting fill level.
• To establish a framework for optimizing formulation development and process understanding.

Main Methods:

• Utilized three AI/ML tools: Random Forest Regression, a symbolic regression tool for reduced-order modeling, and Artificial Neural Networks (ANN).
• Trained models using bulk powder characteristics and processing parameters (RPM, Mixing Blade size, Wall Friction Angle, Feed Rate).
• Validated models on single-component mixtures and a four-component paracetamol formulation.

Main Results:

• ANN demonstrated the highest accuracy in fill level prediction (r² = 0.97).
• Ranked feature importance: RPM, Mixing Blade size, Wall Friction Angle, Feed Rate.
• ANN predicted fill level for a paracetamol formulation with a mean absolute error of 1.4% across different RPMs.

Conclusions:

• AI/ML models provide a robust framework for predicting blender fill levels and understanding process-formulation interactions.
• This approach enables a 'first-time-right' strategy for formulation development, reducing experimental needs.
• The study enhances risk assessment and holistic understanding of the processing environment for powder formulations.