(71a) Surrogate Approaches for Modeling Pharmaceutical Flowsheet Models

This presentation discusses our approach in end-to-end modeling of Pfizer’s continuous powder mixing process (Blackwood et al., 2019), an important step in the production of solid oral dosage forms. By leveraging various data-driven frameworks and tools for sampling, model selection, and steady-state and dynamic modeling, we effectively develop generalizable data-driven surrogate models for the process’ flowsheet model. In this presentation, we will discuss our end-to-end modeling approach, ranging from feature selection and analysis, active learning techniques to generate high-quality data, our model recommender and build tool for steady-state modeling, and a recurrent neural network architecture for dynamic modeling. Various critical quality attributes of the tablets were modeled via our surrogates. The trained surrogate models were generalizable across the input domain, computationally cheaper and faster in predicting outputs than first-principles-based flowsheet models, and accurate in predicting steady-state behavior and dynamic profiles of time-varying responses for fluctuations in different inputs.

We extended the implementation of our frameworks to model a solvent swap pharmaceutical process (Papadakis et al., 2016), where an existing solvent is substituted by a new solvent to accomplish a different task along a production line. Typically, liquid-liquid extraction and batch distillation are used to replace the original solvent by a swap solvent. To model this process, our adaptive sampling algorithm generated sufficient data at optimal locations of the domain, while our model selection and training tools were used to develop accurate surrogates for the composition of the original and swap solvent in the final process streams exiting a batch distillation unit. Our surrogates were accurate in capturing the steady-state and dynamic behavior of solvent composition in the batch still for varying flowsheet inputs.

The different surrogate-modeling functionalities have been embedded within a single tool, capable of modeling any pharmaceutical flowsheet model developed in gPROMS Formulated Products or Aspen Plus environment. Our tool can play a key role in expediting digitalization of industrial and pharma processes.

References:

Blackwood, D.O., Bonnassieux, A., Cogoni, G., 2019. CONTINUOUS DIRECT COMPRESSION USING PORTABLE CONTINUOUS MINIATURE MODULAR & MANUFACTURING (PCM&M), in: Ende, M.T., Ende, D.J. (Eds.), Chemical Engineering in the Pharmaceutical Industry. Wiley, pp. 547–560. https://doi.org/10.1002/9781119600800.ch72

Papadakis, E., Tula, A.K., Gani, R., 2016. Solvent selection methodology for pharmaceutical processes: Solvent swap. Chem. Eng. Res. Des. 115, 443–461. https://doi.org/10.1016/j.cherd.2016.09.004