Closed-loop formulation–process optimization under unknown feasibility for extended-release pellets

Pharmaceutical formulation development frequently encounters formulation–process conditions that fail before a viable product is produced, leaving downstream performance unobservable and experimental feasibility unknown a priori. Here, we demonstrate a closed-loop formulation-development workflow that jointly optimizes formulation and process variables while learning experimental feasibility from physical outcomes. Using caffeine-loaded extended-release pellets produced by wet granulation and extrusion-spheronization, ANDROMEDA 1 selected formulation compositions and processing conditions within a large mixed categorical and numerical design space. Across 51 experiments performed in 13 sequential batches, 29 conditions produced pellets suitable for downstream characterization and 22 (43%) were infeasible. Failed experiments were retained as feasibility observations without assigning artificial drug release scores. Among feasible formulations, the best observed deviation from a predefined four-hour zero-order release target decreased from 1.63 to 0.17 fraction-hours. Release performance, pellet morphology and drug loading followed distinct trajectories, with an empirical trade-off between release deviation and median pellet aspect ratio. Retrospective Gaussian-process modelling showed that replacing an excipient-slot encoding with slot-independent material mass fractions increased leave-one-out cross validated R2 for release prediction from 0.39 to 0.60. Adding four derived wet-mass descriptors reduced release prediction R2 to 0.38 but yielded the highest morphology prediction R2 (0.64). Neither alternative representation improved forward-in-time release prediction. Together, these results show that unknown feasibility, competing product attributes and formulation representation can be incorporated within a single adaptive formulation-process workflow. Analyses further identify design space coverage and representation choice as important considerations for machine learning-guided pharmaceutical development.