A Carrier-Limited Release Ceiling in Electrospun Nanofiber Drug Delivery for Regenerative Endodontics: Machine Learning on a Twenty-Five-Run Experimental Dataset

Regenerative endodontics needs an intracanal medicament that disinfects a necrotic root canal for a one- to four-week dressing interval without killing the stem cells that must then regenerate the pulp. Electrospun antibiotic-loaded nanofibers are the leading candidate, and formulation development proceeds one mat at a time. We ask what machine learning can extract from a single completed 25-run electrospinning campaign, and reach a conclusion that no individual experiment in that campaign could support. We build a three-stage framework on an endodontic study of poly(vinylpyrrolidone) (PVP) fibers loaded with metronidazole, ciprofloxacin, minocycline and Ca(OH)2. Stage 1 learns composition → morphology: nanoscale spinnability is predicted at 96.0% leave-one-out accuracy (24/25, versus an 80.0% majority baseline), with the single error at the Ca(OH)2 4%→5% failure boundary, and mean fiber diameter to 22.3% MAPE (R2 = 0.47), though per-system errors still range from 51 nm where the space is densely sampled to 515 nm where it is not. Stage 2 transfers a release model from an independent literature corpus, because the source study measured no release kinetics. Here the framework’s most useful output is a rejection: carrier class alone separates release regimes perfectly (median time to 90% release 2.0 h for hydrophilic versus 192 h for hydrophobic carriers, Mann–Whitney U = 0, p < 10−4), carrier identity explains nearly all cross-study variance in log10 t90 (R2 = 0.45 alone versus 0.50 with diameter and loading added), and diameter has no significant within-class effect (ρ = 0.51, p = 0.052). No formulation in the entire measured design space—20 spinnable mats, all PVP—can reach the clinical window; the shortfall is a factor of nine beyond the corpus’s longest hydrophilic release. Stage 3 finds three Pareto optimal formulations over the measured antimicrobial, cytocompatibility and dentin-integrity objectives, and shows that fiber formats trade a 4× potency loss for a strict cytocompatibility gain over pastes. We argue this is the shape of useful ML for small experimental campaigns: not a better interpolator, but an instrument that reads a design-space ceiling off data the experimentalist already owns.