A tutorial overview of model predictive control for continuous crystallization: current possibilities and future perspectives

This paper presents a systematic approach to the advanced control of continuous crystallization processes using model predictive control. We provide a tutorial introduction to controlling complex particle size distributions by integrating population balance equations with detailed models of various continuous crystallizers. Since these high-fidelity models are often too complex for online optimization, we propose the use of data-driven surrogate models that enable efficient optimization-based control. Through two case studies, one with a low-complexity system allowing direct comparison with traditional methods and another involving a spatially distributed crystallizer, we demonstrate how our approach enables real-time model predictive control while maintaining accuracy. The presented methodology facilitates the use of complex models in a model-based control framework, allowing precise control of key particle size distribution characteristics, such as the median particle size $d_{50}$ and the width $d_{90} - d_{10}$. This addresses a critical challenge in pharmaceutical and fine chemical manufacturing, where product quality depends on tight control of particle characteristics.

💡 Research Summary

This tutorial paper provides a comprehensive roadmap for implementing model predictive control (MPC) in continuous crystallization processes, with a focus on controlling the full particle size distribution (PSD). The authors begin by highlighting the importance of precise PSD control in pharmaceutical and fine‑chemical manufacturing, where product efficacy and regulatory compliance depend on tight size specifications. While continuous crystallizers offer advantages over batch operations—such as reduced variability, lower capital cost, and improved scalability—real‑time advanced control remains challenging because accurate models involve population balance equations (PBEs) coupled with spatially distributed fluid dynamics, resulting in high‑dimensional partial differential equations (PDEs).

The literature review identifies a gap: existing works either address PBE solution methods, spatial crystallizer modeling, surrogate‑based machine learning, or MPC, but none integrate all four aspects. To fill this gap, the paper is organized into a left‑to‑right workflow (Fig. 1). Sections 3 and 4 describe high‑fidelity modeling: (i) various numerical PBE solvers (finite‑volume, orthogonal collocation on finite elements, quadrature method of moments, GPU‑accelerated schemes) and (ii) detailed continuous‑phase models for different crystallizer architectures (well‑mixed MSMPR, plug‑flow, slug‑flow, oscillatory‑baffled). The resulting models capture nucleation, growth, agglomeration, and breakage across both particle‑size and spatial dimensions.

Because such detailed models are far too computationally intensive for online optimization, Sections 5 and 6 introduce data‑driven surrogate modeling. Large‑scale simulation data are used to train neural networks (feed‑forward, recurrent, auto‑encoders), Gaussian process regressors, and sparse polynomial chaos expansions. The surrogate models are constructed to be differentiable (using PyTorch/JAX) so that they can be embedded directly into the MPC prediction horizon. Uncertainty quantification (Bayesian NNs, ensembles) and model robustness techniques are also discussed, ensuring that the controller can tolerate sensor noise and model mismatch.

The MPC formulation incorporates multiple objectives—maintaining a target median size d₅₀, minimizing the width d₉₀ − d₁₀, and limiting energy consumption—while respecting constraints on concentrations, temperatures, and equipment limits. The surrogate‑based prediction model enables the use of fast nonlinear solvers (sequential quadratic programming, ADMM) within the receding horizon, achieving sub‑10 ms computation times. Two case studies validate the approach. The first, a well‑mixed MSMPR crystallizer, allows direct comparison with traditional moment‑based MPC, demonstrating that the surrogate‑based controller achieves comparable control performance with a 20‑fold reduction in computational load. The second case involves a spatially distributed plug‑flow crystallizer; the surrogate trained on 1‑D spatial‑1‑D size PBE data successfully tracks d₅₀ and distribution width under rapid feed‑rate and temperature disturbances, with control updates faster than 0.01 s.

The paper concludes by outlining future research directions: extending the framework to multi‑component, multi‑reaction systems; implementing online surrogate updating with streaming process data; integrating digital twins and cloud‑based optimization for large‑scale plants; and developing standardized APIs for seamless deployment in pilot‑scale facilities. All code (PBE solvers, surrogate training pipelines, MPC implementation) and an interactive Streamlit dashboard are released on GitHub, encouraging reproducibility and community development.

In summary, the authors demonstrate that high‑fidelity PBE‑based models can be compressed into efficient, differentiable surrogates, enabling real‑time MPC that directly regulates the full PSD in both well‑mixed and spatially distributed continuous crystallizers. This bridges a critical gap between detailed mechanistic modeling and practical advanced control, paving the way for tighter product quality control in the pharmaceutical and fine‑chemical industries.