Bayesian Two-Process Weibull Modelling of PCL-Blend-PEG Coating Degradation in Simulated Body Fluid:...

Predicting how polycaprolactone-poly(ethylene glycol) (PCL-blend-PEG) implant coatings lose mass in simulated body fluid (SBF) would shorten costly 16-week immersion campaigns, but published studies typically provide only three to five compositions, too few for flexible data-driven models. We re-analysed the 16-week mass-loss curves of dip-coated PCL, PCL-blend-PEG 3:1, PCL-blend-PEG 1:3, and PEG films on titanium using a fully specified Bayesian workflow. A mechanistically motivated two-process Weibull model, combining a fast PEG dissolution term and a slow PCL hydrolysis term with composition-dependent time scales, was compared against a parsimonious single effective-Weibull baseline under a prespecified decision rule with exact leave-one-composition-out (LOCO) and future-time refits. Both models converged cleanly (maximum R^≤1.005, zero divergences), yet the two-process model did not outperform the baseline: pooled leave-one-composition-out mean absolute error was 43.7 percentage points for the two-process model versus 37.6 for the baseline, and pooled temporal error 13.2 versus 11.8 percentage points, respectively. The PEG parameters of the two-process model were strongly correlated (r=0.956), precluding separate identification of PEG and PCL kinetics. Empirical coverage of the nominal 90% LOCO intervals was only 21.4% for the two-process model and 3.6% for the baseline. Composition-wise, residuals localise the failure: PEG release inside blends is matrix-retarded rather than instantaneous. We report the negative result transparently, quantify the additional data needed, and release the full reproducible pipeline. These findings define the current predictive limits imposed by the available experimental data and provide guidance for future experimental design.