Machine learning assisted in-line calibration models for near-infrared spectroscopy in dry granulation

Dry granulation is a critical pharmaceutical manufacturing process employed for solid oral dosage forms. The process accommodates high throughput and variable particle properties. However, this makes controlling and monitoring the process essential. Near-infrared (NIR) Spectroscopy is a widely used characterization tool in pharmaceutical manufacturing processes due to its fast and non-invasive nature. In this study, we developed an in-line NIR spectroscopy system with a specialized sampling device (SpiderWheel) for continuous granular flow monitoring. We compared nonlinear calibration techniques, including artificial neural networks (ANN), Bayesian neural networks via Monte Carlo Dropout (BNN-MCD), and least-squares support vector machines (LS-SVM), using partial least squares (PLS) regression as a linear baseline. Various spectral preprocessing methods were evaluated to improve model performance. The BNN-MCD model showed superior predictive accuracy for in-line measurements of granule size distribution (R² = 0.993, RMSPE = 5.72%). It also accurately predicted other granular physical attributes, including porosity (R² = 0.984), specific pore volume (R² = 0.984), and envelope density (R² = 0.980). Optimization with Savitzky-Golay filtering and standard scaling significantly improved nonlinear model performance. This study advances Process Analytical Technology (PAT) by developing accurate, machine learning assisted NIR spectroscopy calibration models for continuous pharmaceutical manufacturing. The proposed system enables precise and real-time characterization of granular properties, supporting Quality by Design (QbD) principles and reliable process monitoring.