Machine Learning Model for the Prediction of Blend Uniformity in Pharmaceutical Powder Blending Process

Purpose: In this study, a data-driven framework for the prediction of blend uniformity in a Y-Cone batch blending process is presented utilizing ML techniques. The aim was to develop a prediction model utilizing PCA-guided feature analysis, identifying the most influential material and process attributes contributing to blend uniformity, where the purpose of the analysis was to interpret feature contributions to principal components, ensuring that variables with meaningful structural influence were highlighted for model understanding and process control design. This approach allowed for the investigation of the factor loadings of each feature. Methods: Key features were selected using multivariate regression analysis and PCA, capturing material variability, density & flow behavior, moisture sensitivity, and process dynamics. A five-year data history containing 7,000 instances was collected from Addis Pharmaceutical Factory (APF) and was preprocessed. To establish a robust predictive relationship, three machine Learning (ML) techniques, namely Feed-Forward Neural Network (FFNN), Random Forest (RF), and extreme gradient boosting (XGBoost), were developed and evaluated. With Mean Squared Error (MSE) as the loss function, the FFNN was trained and validated using the Adam optimization algorithm, incorporating dropout regularization (0.1) and early stopping to prevent overfitting. While the XGBoost model was optimized through gradient boosting with an MSE objective function and regularization terms to control model complexity, BO was employed for the RF model hyperparameter tuning, with bagging and random feature subset selection. Results: Model’s performance was evaluated using the RMSE, MAE, and R² across the split datasets. Accordingly, FFNN attained an R² value of 0.9919, demonstrating higher predictive capability, outperforming the other two models and models from the literature, demonstrating superior capability in capturing nonlinear blending dynamics. The FFNN also exhibited the lowest Root Mean Squared Error (RMSE) (0.0018) and Mean Absolute Error (MAE) (0.0012), confirming its robustness and generalization ability. Conclusion: These findings demonstrate the suitability of neural network-based approaches in capturing the nonlinear relationships governing pharmaceutical blend uniformity and support their application as predictive tools for process optimization and control.