(579d) Dynamic Model Based Design of Experiments for Hybrid Modelling of Bioprocess

One of the main hurdles in the computer-aided design and optimization of industrial bioprocesses is the limited capability of models to accurately represent biological systems. Mechanistic models face challenges in defining and integrating every relevant biomechanism, whereas data-driven models fall short in extending predictions beyond the existing data distribution. Hybrid models stand as a potential solution, embodying the strengths of both approaches [1]. Unfortunately, building a hybrid model that can accurately balance its data-driven and first principles components is not trivial. Additionally, ongoing research is focused on the assessment and mitigation of prediction uncertainty within the hybrid models.

This study presents a new approach to building hybrid models, combining a mechanistic foundation informed by existing knowledge of biological mechanisms with data-driven elements. These data-driven elements are integrated using statistical model selection techniques, optimizing both fitting precision and the capacity for extrapolation [2]. Statistical model selection methods such as the Bayesian Information Criterion (BIC) [2], the Akaike Information Criterion (AIC) [3], and the corrected Akaike Information Criterion (AICc) [4] are embedded into the framework. All three criteria are used as an ensemble to find candidate hybrid model structures. Then, to find the optimum hybrid model structure, these candidate structures are further discriminated through dynamic MBDoE loops [5, 6]. At the end, the parameter uncertainty is quantified and further reduced through the optimal experimental design following the D-optimal criterion [5, 7] to enhance the overall robustness of the hybrid model.

The proposed methodology is tested on a fed batch microalgae cultivation case study for biodiesel production. The mechanistic structure is formulated on the mass balance of state variables. Factors such as light attenuation, substrate, and nitrate consumption and their impact are, as the data-driven component, represented by polynomial regression and artificial neural network (ANN) . The optimal structure and order of the polynomial equation and ANN are determined through three rounds of the proposed MBDoE framework. After that, parameter uncertainty is further reduced with the optimal experimental design. The results show that the proposed framework can construct a statistically sound hybrid model with limited data for bioprocess modelling. Furthermore, the measurement and minimization of model uncertainty exhibits significant promise in the enhancement of hybrid model-based process control and optimization.

References

[1] M. Von Stosch, R. Oliveira, J. Peres, and S. F. de Azevedo, "Hybrid semi-parametric modeling in process systems engineering: Past, present and future," Computers & Chemical Engineering, vol. 60, pp. 86-101, 2014.

[2] A. A. Neath and J. E. Cavanaugh, "The Bayesian information criterion: background, derivation, and applications," Wiley Interdisciplinary Reviews: Computational Statistics, vol. 4, no. 2, pp. 199-203, 2012.

[3] Y. Sakamoto, M. Ishiguro, and G. Kitagawa, "Akaike information criterion statistics," Dordrecht, The Netherlands: D. Reidel, vol. 81, no. 10.5555, p. 26853, 1986.

[4] C. M. Hurvich and C. L. Tsai, "A corrected Akaike information criterion for vector autoregressive model selection," Journal of time series analysis, vol. 14, no. 3, pp. 271-279, 1993.

[5] G. Franceschini and S. Macchietto, "Model-based design of experiments for parameter precision: State of the art," Chemical Engineering Science, vol. 63, no. 19, pp. 4846-4872, 2008.

[6] I. Y. Tumer, C. Hoyle, D. C. Jensen, and S. Hunter, "Validating model-based design simulation: The impact of abstraction and fidelity levels," in 2015 International Conference on Complex Systems Engineering (ICCSE), 9-11 Nov. 2015 2015, pp. 1-6, doi: 10.1109/ComplexSys.2015.7385994.

[7] W. G. Hunter and A. M. Reiner, "Designs for Discriminating between Two Rival Models," Technometrics, vol. 7, no. 3, pp. 307-323, 1965, doi: 10.2307/1266591.