Run-to-Run-Based Model Predictive Control of Protein Crystal Shape in Batch Crystallization

Joseph Sang‐Il Kwon, Michael Nayhouse, Gerassimos Orkoulas, Dong Ni, Panagiotis D. ChristofidesIndustrial 2014 · DOI

TL;DR

This paper introduces a novel "run-to-run" model predictive control (R2R-MPC) strategy to precisely control the shape of protein crystals during batch crystallization. By learning from previous batch runs and adapting its internal model, the R2R-MPC significantly improves control performance and achieves desired crystal shapes much faster than conventional methods, even in the presence of process drift and variability.

What Problem Does This Paper Address?

Producing crystals with specific, desired shapes and sizes is crucial in industries like pharmaceuticals, as these properties directly influence drug efficacy, stability, and dissolution rates. Batch crystallization processes, however, are notoriously difficult to control precisely due to inherent variability in factors like solubility and crystal growth rates, as well as gradual "drift" in these parameters over repeated batches. Traditional model predictive control (MPC) systems, which rely on a fixed process model, struggle with this variability and drift, leading to inconsistent product quality and slow convergence to desired outcomes. The paper aims to bridge this gap by developing a control strategy that can adapt to these changing conditions and consistently produce crystals with the target shape.

Methods

The core of the proposed solution is a Run-to-Run Model Predictive Control (R2R-MPC) framework. This approach combines the predictive power of MPC with a learning mechanism that utilizes data from previous batch runs. Here's a breakdown of the key components:

  • Conventional MPC Foundation: The system starts with a standard MPC that uses a nominal model of the crystallization process (derived from moment models of a population balance equation, and mass/energy balances) to compute optimal jacket temperatures for crystal growth.
  • Kinetic Monte Carlo (kMC) Simulation: A detailed kMC model is used to simulate the crystal growth process, providing a realistic platform for testing and validating the control strategies. This model captures the complex surface processes involved in crystal formation.
  • Batch-to-Batch Parameter Estimation: This is the "run-to-run" learning aspect. After each batch concludes, post-batch measurements (crystal shape, size, and number) are used to solve a multivariable optimization problem (MOP). This MOP estimates "correction factors" that adjust the parameters of the MPC's internal model (specifically, solubility and crystal growth rates) to account for the observed drift and variations from the previous run.
  • Exponentially Weighted Moving Average (EWMA): To handle any remaining offset in crystal shape that the parameter estimation might not fully capture, an EWMA scheme is employed. This smooths out short-term fluctuations and helps maintain stability.
  • Integrated Approach: The R2R-MPC integrates these components: it uses the adapted model parameters from the MOP and the EWMA correction to compute optimal jacket temperatures for the next batch. This iterative learning process allows the controller to adapt and improve its performance over successive batches.

Datasets

The paper does not explicitly mention using external datasets for training. Instead, it relies on:

  • Simulated Data: The primary "dataset" for evaluation comes from extensive simulations using a Kinetic Monte Carlo (kMC) model of tetragonal hen egg white (HEW) lysozyme crystallization. This simulation environment is designed to mimic real-world process drift and variability by perturbing parameters like pH and introducing gradual changes in solubility and growth rates.
  • Literature Data: The parameters for the kMC model, nucleation rates, and solubility equations are calibrated using experimental data from existing literature (references [20]-[26]).

Key Insights and Ideas

  • Adaptive Control is Crucial for Batch Processes: The study powerfully demonstrates that static models are insufficient for batch crystallization. The inherent drift and variability necessitate an adaptive control strategy that learns and adjusts over time.
  • Combining Parameter Adaptation and Offset Correction is Key: The R2R-MPC's success stems from its dual approach: adapting the core process model parameters (Parameter Adaptive Control - PAC) and using EWMA for residual offset cancellation (Offset Drift Cancellation - ODC). This combination tackles both systematic drift and transient variations effectively.
  • Faster Convergence to Desired Crystal Shapes: The R2R-MPC achieves the target crystal shape distribution significantly faster (3 batches) compared to conventional methods (24 batches for EWMA-type CSC). This is a major practical advantage for industrial production.
  • Robustness to Different Drift Types: The R2R-MPC shows good performance even when subjected to different types of process drift (decaying, constant, and random), highlighting its practical applicability.

Key Results

  • Rapid Convergence: The proposed R2R-based MPC successfully produced crystals with a desired shape distribution after just three batch runs.
  • Superior Performance vs. Conventional Methods: In contrast, an EWMA-type constant supersaturation control (CSC) strategy required 24 batch runs to achieve the same objective. A conventional MPC without R2R adaptation showed progressively worse performance over repeated runs.
  • Accurate Solubility Prediction: The R2R-MPC's predicted solubility closely matched the true solubility, with an offset of approximately 2%.
  • Effective Handling of Process Drift: The controller effectively suppressed inherent variations and process drift in solubility and crystal growth rates, leading to consistent crystal shape control.
  • Impact of Learning Factor (λ): While varying the learning factor (λ) did not significantly alter the convergence speed for the R2R-MPC, smaller λ values were found to be beneficial for reducing the impact of measurement noise without sacrificing convergence.

Limitations and Future Work

  • Linear Parameterization: The paper acknowledges that while it addresses certain nonlinear parametric uncertainties by re-parameterization, handling more complex nonlinearities might require solving constrained nonlinear optimization problems, which can be computationally intensive.
  • Measurement Noise: While a small learning factor (λ) helps mitigate measurement noise, the paper notes that for very high noise levels, R2R controllers might struggle if past measurements are not well-correlated with future runs. A low-pass filter is suggested as a potential enhancement.
  • Unmodeled Nucleation Uncertainty: The study showed that while the R2R-MPC could compensate for unmodeled nucleation rate uncertainty, the control performance was slightly degraded. The authors suggest that a nonlinear ODC scheme could improve robustness in such cases.
  • Future Work: The authors hint at the potential for adapting the learning factor (λ) dynamically based on the observed process drift rate to further improve performance in rapidly changing environments. Exploring nonlinear ODC schemes for better handling of unmodeled uncertainties is also a promising avenue.

Summary updated

Abstract

In this work, we develop a novel run-to-run-based model predictive controller (R2R-based MPC) for a batch crystallization process with process drift and inherent variation in solubility and crystal growth rates. In order to achieve the production of crystals with desired product qualities, a conventional MPC system with nominal process model parameters is initially applied to a batch protein crystallization process. However, the mismatch between the process model and the actual process dynamic behavior because of the process drift and variability becomes severe as batch runs are repeated. To deal with this problem of batch-to-batch variability, after each batch is over, the post-batch crystal attribute measurements, including average crystal shape and size and the number of crystals, are used to estimate off-line the drift of the process model (used in the MPC) parameters from nominal values via a multivariable optimization problem. Along with the adapted controller model parameters, the exponentially weighted moving average (EWMA) scheme is used to deal with the remaining offset in the crystal shape values and thereby to compute a set of optimal jacket temperatures. Furthermore, the crystal growth in the batch crystallization process is modeled through kinetic Monte Carlo simulations, which are then used to demonstrate the capability of the proposed R2R-based MPC scheme in suppressing the inherent variation and process drift in solubility and crystal growth rates. It is demonstrated that the production of crystals with a desired shape distribution is successfully achieved after three batch runs through the use of the proposed R2R-based MPC, while it takes 24 batch runs for the system with the EWMA-type constant supersaturation control to achieve the same objective.
BibTeX
@inproceedings{kwon2014run,
  author = {Joseph Sang‐Il Kwon and Michael Nayhouse and Gerassimos Orkoulas and Dong Ni and Panagiotis D. Christofides},
  title = {{Run-to-Run-Based Model Predictive Control of Protein Crystal Shape in Batch Crystallization}},
  booktitle = {Industrial},
  year = {2014},
  doi = {10.1021/ie502377a},
}
Figure