An efficient model development strategy for bioprocesses based on neural networks in macroscopic balances

Area of Science:

• Biochemical Engineering
• Process Modeling
• Computational Chemistry

Background:

• Macroscopic balances often contain inaccurately known terms.
• Neural networks offer powerful tools for modeling complex relationships.
• Existing modeling strategies include black box and white box approaches.

Purpose of the Study:

• To introduce and validate the serial gray box modeling strategy for biochemical conversions.
• To demonstrate the extrapolation capabilities of the serial gray box model.
• To compare the serial gray box strategy with black box and white box methods.

Main Methods:

• Combining macroscopic balances with neural networks.
• Utilizing identification data covering the input-output space of inaccurately known terms.
• Applying the strategy to model enzymatic batch conversion of penicillin G.

Main Results:

• The serial gray box strategy allows reliable extrapolation using accurately known terms.
• Models developed using this strategy exhibit superior extrapolation properties compared to black box models.
• The strategy proved successful in modeling penicillin G enzymatic conversion with real-time results.

Conclusions:

• Serial gray box modeling provides reliable extrapolation and wider applicability than black box models.
• It offers a shorter development time than white box models when detailed system knowledge is limited.
• This hybrid approach balances data-driven and knowledge-driven modeling effectively.