Scale-independent prediction of ribbon and granule quality in dry granulation process based on transfer learning for...

In dry granulation, differences in equipment geometry, compaction mechanisms, and powder flow behavior prevent quality prediction models based on a single device from achieving cross-scale and cross-equipment predictions of ribbon solid fraction (SF) and granule quality. Focusing on ten traditional Chinese medicine "Jichuan decoction" extracts, this study proposes a hybrid transfer learning (TL) framework to perform two-level device transfer for scale-independent prediction of ribbon and granule quality. In the first transfer level, a source domain dataset is constructed using uniaxial compaction experiments, and a target domain dataset using roller compaction experiments; material compression characteristics are extracted via the Johanson model. T-shaped partial least squares regression (T-PLSR) is then introduced for feature alignment and knowledge transfer. In the second transfer level, a Johanson + T-PLSR hybrid model is built, using the compression parameters output by the physical model as input features to jointly predict ribbon solid fraction and density, granule yield, and friability. This hybrid model overcomes the limitation of the Johanson model, which only predicts ribbon solid fraction, achieving end-to-end prediction from material compaction behavior to final granule quality. Validation results show that the prediction accuracy of the hybrid transfer model is significantly superior to pure physical or pure data-driven models, while the inclusion of physical parameters enhances model interpretability and generalization ability.