Abstract
Continuous pharmaceutical manufacturing (CM) offers improved quality assurance, operational agility, and supply resilience, yet process development remains dominated by expensive trial-and-error experimentation and high-dimensional space exploration. Motivated by ICH Q13, we develop a generative inverse-design framework that maps target product quality to feasible process recipes for an integrated twin-screw wet granulation and segmented fluidized-bed drying line. The framework integrates three components: (i) a Conditional Variational Autoencoder (CVAE) generator that proposes process parameter sets conditioned on desired Critical Quality Attributes (CQAs), (ii) a Gaussian Process (GP) surrogate validator that screens candidates for manufacturing feasibility, and (iii) SHapley Additive exPlanations (SHAP) to interpret the generated designs. Training data were produced from a validated gPROMS digital twin of the Diamond Pilot Plant (DiPP) ConsiGma-25 line, covering liquid -to-solid ratio, drying temperature, drying time and air flowrate, with CQAs including granule moisture content, average particle size and porosity. The trained CVAE generated ~50, 000 candidate recipes and learned constrained feasible regions of the design space. Across seven operating scenarios, generated recipes achieved low deviations from targets. For a target-center case, deviations were 2.0% (moisture), 0.4% (particle size) and 0.5% (porosity). Edge cases remained acceptable, with the largest deviation observed for moisture in a high-moisture scenario (8.8%). SHAP analysis highlighted drying time and liquid-to-solid ratio as the dominant drivers of moisture, while liquid-to-solid ratio governed particle size and porosity. Overall, the approach enables rapid, explainable exploration of CM design spaces, reducing experimental burden and supporting QbD-aligned development for integrated continuous processes.