By Dr. Zoltan K. Nagy, Purdue University

This talk presents resource-efficient, uncertainty-aware digital design workflows that combine targeted and automated experimentation with mechanistic and data-driven models to enable robust design of batch and continuous crystallization systems. The first part introduces a mechanistic population balance model (PBM)-driven framework with quantified prediction uncertainty, establishing a probabilistic model validation approach. A key challenge in PBM development is identifying dominant crystallization mechanisms and appropriate kinetic models. To address this, a systematic kinetic model development workflow is proposed, followed by model-based design of experiments (mb-DoE) to generate informative data and improve model accuracy. These methods were demonstrated on a modular autonomous crystallization system (MACS) in recycle operation, achieving substantial savings in material and time. In parallel, Bayesian optimization (BO) algorithms were integrated to accelerate nonlinear optimization in parameter estimation and mb-DoE, while supporting likelihood-based uncertainty characterization. Building on these foundations, stochastic optimization formulations incorporate model uncertainty directly into operating policy design. A surrogate-assisted sample-average approximation approach enables efficient solution, yielding strategies robust to both parametric and operational variability. Recognizing that mechanistic PBMs may be impractical in some settings – due to resource limitations or the complexity of critical quality attributes such as crystallinity -complementary data-driven workflows are introduced. Active learning and statistical machine learning models guide experimentation for design space exploration, while synthetic and experimental data augmentation enhance predictive accuracy under limited budgets. Machine learning surrogates are then used for design space analysis and optimization, and closed-loop BO workflows are explored to iteratively drive processes toward optimal performance. Together, these contributions establish end-to-end digital design strategies – mechanistic where models are tractable, and data-driven where they are not – that reduce experimental burden, explicitly account for uncertainty, and accelerate the deployment of digital tools in pharmaceutical crystallization.