Hybrid models are being applied increasingly to fermentation processes where they blend the benefits of mechanistic and data-driven modeling while minimizing their limitations. They also are faster to build and appear to be more accurate, thus accelerating fermentation process scaleup.
The key benefit may be the ability of hybrid models to predict changes accurately in complex situations in which knowledge is incomplete. Because they can reduce process rework, scaleup is streamlined and deployment is accelerated, according to a review by Krist Gernaey, PhD, professor of industrial fermentation technology, Technical University of Denmark (DTU), and colleagues from DTU and Novonesis.
To reach this conclusion, Gernaey’s team surveyed relevant literature.
“One of the key papers [by Harini Narayanan and colleagues at ETH Zurich and DataHow]…highlighted the differences in data requirements and extrapolation capabilities of data-driven, hybrid, and mechanistic models,” Gernaey says.
As first author Mariana Albino, tells GEN, Narayanan’s work showed that, “In an interpolation scenario, all models performed satisfactorily regarding the obtained mean squared error (MSE). The main difference…was the number of training runs required. The data-driven and the mechanistic models required, respectively, the most (50 runs) and the least (10 runs) amount of data. The mechanistic model achieved a higher MSE.”
To extrapolate, use hybrid models
Combining historical data with the ability to extrapolate data from related situations—hybrid models—significantly enhanced successive models’ value. Adding data for the rate of accumulation, mass balances, or specific rates allowed the models to be optimized in only 30 training runs, and further lowered the MSE. In contrast, models that included either kinetic terms or the specific growth and death rate required 50 runs to achieve that same MSE.
When the models were required to extrapolate data, the differences became more notable.
“Even with 50 training runs, the data-driven model did not yield a satisfactory root MSE,” Albino says. [The data-driven model’s] extrapolation capabilities are poor, limiting its value for control and optimization activities. The best performance,” she concludes, “was once again achieved by one of the hybrid models.”
The mechanistic model performed well and required less training data. However, she cautions, “Mechanistic models can be time-consuming to develop. They require extensive validation work and intensive process knowledge.”
Gernaey adds, “In many cases, they cannot cope with situations such as a change in biomass or product yield throughout a fermentation run.”
In contrast, “hybrid modes… can cope with some of the limitations of a mechanistic model,” he continues, thanks to their data-driven elements. “They are especially useful in situations where a process (such as filamentous fermentation) undergoes slow changes.”
Hybrid models excelled at extrapolation, the researchers explain, “making them suitable for process control outside the tested process conditions.”
Ultimately, hybrid models offer a complementary alternative to either purely data-driven or mechanistic models. The increasing numbers of case studies highlighting the benefits of hybrid models should spur the development of industrially relevant pilot studies, Gernaey predicts.
