Filament-level programming and machine learning enable tunable release kinetics beyond material limits

Much as a prism resolves white light into distinct wavelengths, filament-level structural programming can resolve a single formulation into distinct drug-release kinetics. Here, we show that release behavior can be extensively programmed within a chemically fixed, nonswelling matrix by organizing drug-loaded, drug-free, and barrier filaments into designed topologies that regulate diffusion accessibility, pathway continuity, and phase-specific resistance. To make this design space computable, we developed a multiscale feature engineering strategy that converted anisotropic three-dimensional (3D) architectures into transport-relevant structural fingerprints. This representation enables interpretable machine learning to predict dissolution behavior, generalize to unseen architectures, and support the inverse design of target release profiles. Coupled with physiologically based pharmacokinetic modeling, this workflow further links programmable in vitro release to disease-relevant exposure control, establishing internal architecture as a computable handle for functional mass-transport design.