Product Design for Batch Processes Based on Iterative Learning-Latent Variable Model Inversion (IL-LVMI)

Abstract

Manufacturing chemical products that meet marketing and regulatory demands is critical in batch processes. This is accomplished by creating well-designed input profiles based on process knowledge or historical data. Over the years, many data-driven product design methods have been developed, including latent variable model inversion, which is known for its convenience and efficiency and has been extensively applied to various batch systems. However, even with well-designed input conditions that meet quality requirements, products may still fall short due to the inevitable model mismatch or unmeasurable process disturbances. To address this issue and improve the quality control performance in batch processes, this work proposes an approach that combines iterative learning and latent variable model inversion (IL-LVMI). The latent variable model captures the correlation between input profiles and quality attributes, and model inversion obtains a design space containing a set of input profiles that satisfy the product specifications. The designed input profiles are then reconstructed and implemented in the target batch processes. To increase the accuracy of the obtained product quality, the employed iterative learning algorithm minimizes the deviation between the actual output and the requirements for each iteration. A null space can exist to calculate input increments, and input profiles can be arbitrarily adjusted along the direction of the null space without affecting the convergence performance. By using IL-LVMI, we obtain not only satisfactory product quality but also an updated design space. The methodology proposed in this study was validated through four scenarios by using a continuous stirred tank reactor. The product design results were promising, and an updated design space was visualized by the proposed IL-LVMI approach.