Modeling of bioprocesses via MINLP-based symbolic regression of S-system formalisms

• Published in: Computers & chemical engineering Vol. 170; p. 108108
• Main Authors: Forster, Tim, Vázquez, Daniel, Cruz-Bournazou, Mariano Nicolas, Butté, Alessandro, Guillén-Gosálbez, Gonzalo
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.02.2023
• Subjects: Artificial neural network; Bioprocess; Low sized dataset; MINLP; Optimization; MINLP; Artificial neural network; Low sized dataset; Optimization; Bioprocess
• ISSN: 0098-1354; 1873-4375
• DOI: 10.1016/j.compchemeng.2022.108108

•A method for building models based on an S-system formalism is proposed.•ODE integration is avoided for model training by a two-stage incremental approach.•Models tailorable in complexity by the maximum number of parameters.•Trained closed-form expressions easily interpretable. Mathematical modeling helps guide experiments more effectively, support process monitoring and control tasks, stabilize product quality, increase consumer safety, or ease specific decision-making tasks for subject matter experts. However, constructing accurate process models can be challenging, especially with bioprocesses, due to complex metabolic mechanisms and data scarcity. This work proposes a method for building models combining a mass balance backbone with a canonical kinetic representation, i.e., the S-system formalism. The model structure and parameters that best describe the studied system are automatically identified by solving a mixed-integer nonlinear programming (MINLP) problem. Following an incremental approach, the integration of ordinary differential equations is avoided. Numerical examples show that our method performs similarly to models based on artificial neural networks, outperforming them in some cases while providing an analytical, closed-form model. Such expressions can be more easily interpreted and optimized in existing algebraic modeling systems.