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

Ribbon splitting, a phenomenon that can occur during the roller compaction operation used in dry granulation processes, can lead to compromised granule uniformity, poor tabletability, and ultimately, off-specification tablet production. Despite its importance, ribbon splitting is still difficult to understand and predict because of the intricate and nonlinear relationships between equipment conditions, process variables, and formulation properties. This study presents a machine-learning framework to model and characterize ribbon splitting by leveraging multivariate data collected from experiments. The framework is based on a Gaussian process regression (GPR)-based data augmentation driven neural network with transfer learning. SHapley Additive Explanations (SHAP) analysis is used to enhance model's interpretability, by extracting feature importance and their influence on ribbon splitting. The framework enables reliable prediction of ribbon quality (thickness and density) based on high R² and low MSE/MAE values, determination of feasible regions corroborated through existing studies, and optimal input conditions that ensure product consistency and regulatory compliance. The proposed approach is flexible, scalable, and generalizable across various operating conditions, enabling early fault detection, accelerating process development, providing actionable insights, and establishing the foundation for closed-loop control and QbD implementation in pharmaceutical dry granulation.