Mechanistic and data-driven modelling of powder adhesion

Powder adhesion to punch surfaces (“sticking”) is a persistent, costly problem in pharmaceutical tablet manufacturing. This thesis addresses critical gaps by uniting data-driven prediction, high-fidelity image analysis, and continuum mechanics in a multi-scale framework spanning molecules, micrographs, and process physics.First, the chemical basis of sticking is quantified using molecular descriptors. A data-driven pipeline compresses more than > 1 800 descriptors into a five-component fingerprint that perfectly separates the sticking behavior of 42 benchmark powders, establishing (for the first time) a cheminformatics route to predict tablet manufacturability.Next, a novel rotation-robust feature-extraction architecture is introduced for 834 punch-face micrographs. Equal-area, staggered annular-sector partitions encode axial symmetry and support a topology-aware discrimination strategy derived from observed sticking patterns. Coupled with a semi-guided segmentation network (F1 > 0.97), the method yields spatially resolved sticking maps that decouple where adhesion occurs from how strongly it occurs. Hierarchical statistical analysis identifies material-specific controls: temperature predominantly governs sticking in acetylsalicylic acid, humidity primarily influences mannitol, and all materials exhibit suppressed sticking near the punch edge.Lastly, thermo-mechanically coupled finite-element simulations elucidate underlying mechanisms. Die-wall friction induces radial stress and density gradients; during punch retraction these gradients drive mixed-mode interfacial separation, concentrating tensile stresses away from the edge—the same locations where high sticking probability is observed. A multilayer-perceptron surrogate trained on simulated fields predicts local sticking with high fidelity, linking material attributes, process conditions, and observed defects.Together, the contributions deliver standardized, spatially explicit metrics for powder adhesion and a rigorous methodology for comparing materials and processes. The framework integrates molecular fingerprints, image-derived sticking maps, and physics-based fields, laying foundations for predictive, mechanistic sticking models and enabling quantitative strategies to mitigate a critical manufacturing loss mode.