Hybrid Models Based on Machine Learning and an Increasing Degree of Process Knowledge: Application to Capture Chromatographic Step

Abstract

In process engineering, two paradigms of modeling approaches exist: the mechanistic and the data-driven approaches with the former being completely based on knowledge while the latter completely based on data. In our previous work, we highlighted the advantages of using hybrid models that explores the synergy between mechanistic and data-driven models. Here we introduce the concept of developing a series of hybrid models constituted by a progressively increasing extent of process knowledge. Thus, aligning the models on the “degrees of hybridization” axis with data-driven model being 0% hybridized and mechanistic model being 100% hybridized. In this work, the proposed concept is demonstrated for the application of a chromatographic capture step where the models are evaluated based on (i) prediction accuracy, (ii) extrapolation ability, (iii) providing process understanding, and (iv) practical application. We show the limitations of both model variant extremes. On one hand, the performance of the mechanistic model is compromised due to an excessive imposition of knowledge, thus affecting its predictive capabilities and efficiency in practical utility. On the other hand, the data-driven model inherently is not suitable for application such as multicolumn chromatography or to gain process understanding. In contrast, a series of hybrid models could be developed with better and versatile performance in term of prediction, extrapolation, process understanding, and practical utility. We show that for general process applications the different hybrid model variants and their ensembles have comparable performance. We illustrate the criteria for selection of a particular hybrid model variant based on different considerations such as complexity of training or model development, acquired understanding, and data requirement.

Model Transferability and Reduced Experimental Burden in Cell Culture Process Development Facilitated by Hybrid Modeling and Intensified Design of Experiments

“…The more knowledge about a certain process is available, the fewer experiments have to be performed when hybrid modeling approaches are chosen to develop a well-performing model ( Yang et al, 2015 ; Bayer et al, 2020c ; Smiatek et al, 2020 ). Recently, the incorporation of mechanistic understanding at different levels in different hybrid modeling approaches was evaluated for chromatography ( Narayanan et al, 2021a ). Moreover, Krippl et al (2020) could demonstrate that by using hybrid modeling, different tangential flow filtration operational modes can be described from the same training data set.…”

Design of Biopharmaceutical Formulations Accelerated by Machine Learning

“…As next steps, further robustness can be added into the surrogate model by incorporating physically relevant constraints (e.g., physiological osmolality) and information gained by experience (Figure E). Such approaches are gaining popularity in bioprocessing under the so-called “hybrid modeling” framework, − which combines knowledge-based and data-driven models. Extending this concept to formulation optimization will be a very interesting direction of future works.…”

Using scientific machine learning to develop universal differential equation for multicomponent adsorption separation systems

“…[7,8] The increasing computing power has allowed researchers to revisit hybrid modelling with a different perspective. For instance, Narayanan et al [9,10] were able to use a hybrid strategy within a system of partial differential equations (PDE) to describe the single-component adsorption in a fixed bed unit. The hybrid strategy provided notable results in modelling the system.…”

“…In this work, we extended the MTI-hybrid strategy to a multicomponent scenario. Furthermore, Narayanan et al [9,10] pointed out that the MTI-hybrid required a pre-identified adsorption model to be used as an initial guess for the hybrid model training.…”

“…It is important to highlight that Narayanan et al [9,10] have proposed other hybrid modelling strategies that do not require a pre-identified isotherm. However, in the present work, the MTI-hybrid model is employed to preserve the ODE physical structure and the linear driving force mass transfer model, which is more faithful to the system-known phenomenology.…”