Iterative model-based optimal experimental design for mixture-process variable models to predict solubility

Crystallization process design relies heavily on predictive solubility models. However, their calibration is resource-and labour-intensive, especially for multicomponent solvent mixtures at different process temperatures. Additionally, solubility data collection often occurs in a constrained design space due to e.g., polymorphism and solvent miscibility limitations. Optimal experimental design techniques enable the efficient use of resources by specifying a (minimum) number of maximally informative experiments focused on improving a statistical criterion for a given model structure in a constrained design space. This work generates D-, A-and I-optimal experimental designs for the commonly applied Van't-Hoff Jouyban-Acree (VH-JA) solubility regression model, in which it is demonstrated that I-optimal designs reduce the experimental burden for model calibration by ap-proximately 25 % as compared to a typical screening dataset. Alternatively, existing da-tasets can be augmented to significantly improve model prediction power. The suggested workflow is applied to two case studies: itraconazole in tetrahydrofuran-water and me-salazine in ethanol-polyethylene glycol-water. The screening datasets of 72 and 212 runs were augmented with 16 additional experiments, resulting in a 33 % and 67 % reduction in the corresponding model prediction variance, respectively, which translates to improved model reliability at unprobed conditions.(c) 2022 Institution of Chemical Engineers. Published by Elsevier Ltd. All rights reserved.

Automated summary

The design of the crystallization process relies on predictive solubility models, but their calibration is resource-intensive. Optimal experimental design techniques can reduce the experimental burden and augment existing datasets to improve model prediction power.