Time-resolved process understanding through hybrid modeling

Common approaches for process modeling

Within the last decades, industries and academia approached the description of bioprocesses differently. While academia was focusing on describing both up and downstream processes mechanistically, the industry built process models mainly on data-driven tools. In both cases, process outputs (measured responses or critical quality attributes) should be described using process parameter inputs. The difference between both approaches is that mechanistic models assume the underlying mechanism of how in- and output are connected. Mainly differential equations are used to describe the process accurately. In the data-driven approach, the underlying mechanism is ignored and a correlation model is created based on input data. While the data-driven approach is much easier to build, the models based on this approach might not be of high quality if there is no true causal relationship between input and output data. Further data-driven models are lacking extrapolation abilities compared to their mechanistic counterparts. On the other side, mechanistic models strongly simplify the true behavior and often are less precise compared to data-driven models within a certain design space. Especially for very complex behaviors like cell growth, valid mechanistic equations can hardly be derived.

A robust and reliable solution for process modeling

That is why we combine the best from both modeling approaches, data, and mechanistic understanding, into one superior model structure, a so-called hybrid model. This kind of model delivers more generic descriptions compared to a purely mechanistic model and also is a well-suited match for extrapolation in contrast to a pure data-driven model.

'Hybrid models bridge the gap between mechanistic understanding and process data'

Typically, in a hybrid model structure, the non-parametric part is represented as a black box while the parametric part is displayed as a white box. These can be in a parallel or a serial structure, as shown in the example below. Herein, the process data is used as input to the black-box model, estimating the specific rates for the white-box model to get a more generic prediction for the model predictions. Such hybrid models can easily be generated using your own process data in our freely available hybrid modeling toolbox, no installation is required.

Figure 1. Serial hybrid model structure

Since process evaluation in the biopharmaceutical industry is a critical task, reliable process models are of high value. While state of the art methods, such as only modeling the process endpoint and time-resolved black-box modeling, do not generate sufficient process understanding for demanded quality by design implementation, hybrid modeling is a suitable candidate to deal with the needs of regulatory authorities.

If these different modeling approaches are compared, the respective limitations and the advantages of using a hybrid model get prominent.

  • Compared to a pure process endpoint model, the entire bioprocess can be predicted in a time-resolved manner using a hybrid model

  • Further, by the incorporation of process knowledge and mechanistic information, the predictions get way more accurate compared to a time-resolved black-box model

Figure 2. Modeling approaches comparison

This proves the hybrid modeling structure to be superior with respect to the model performance, and its usage moreover also enables: