Process analytical technologies provide the opportunity for monitoring and control of crystallization processes. However, the implementation of optimal feedback policies can be challenging due to the lack of accurate models. As an alternative to physically based models, machine learning can overcome certain challenges, such as uncertainty in model structure. However, this approach may suffer from other challenges including robustness to unmodeled dynamics and the lack of sufficient validation data. These challenges will be discussed in the context of several projects on paracetamol crystallization in solvent mixtures of ethanol and water using ReactIR and FBRM for real-time monitoring and control.
