A Hierarchical Machine Learning Surrogate investigates the impact of adverse ribbon properties on drug product quality in a dry granulation process.

Recognizing ribbon splitting as a phenomenon that impacts ribbon quality and can potentially influence granule size distribution and, consequently, tablet quality attributes such as tensile strength, dissolution, and uniformity, this work develops a hierarchical machine learning surrogate to study the effect of process and material conditions in a dry granulation line on granule and tablet properties under ribbon splitting conditions. Ribbon splitting is treated as a latent, process-induced structural state arising from elastic recovery of the compacted ribbon. The surrogate connects ribbon-level physical parameters to granule size distributions and further to tablet characteristics, supported by multi-level experiments. The surrogate not only provides robust predictions of granule and tablet properties but also enables statistical analysis of which downstream drug product attributes are most impacted and their corresponding mechanism. The latter helps in providing process insights essential for Quality by Design (QbD) strategies and real-time control. The hierarchical model is also deployed to illustrate design space regions that can help mitigate the impact of adverse ribbon quality on granule and tablet characteristics across two distinct API formulations.