Modeling biosurfactant production from agroindustrial residues by neural networks and polynomial models adjusted by particle swarm optimization
Biosurfactants are molecules with wide application in several industrial processes. Their production is damaged due to inefficient bioprocessing and expensive substrates. The latest developments of strategies to improve and economize the biosurfactant production process use alternative substrates, optimization techniques, and different scales. This paper presents a study to compare the performances of classical (polynomial models) and modern tools, such as artificial intelligence to aid optimization of the alternative substrate concentration (alternative based on beet peel and glycerol) and process parameters (agitation and aeration). The evaluation was developed in two different scales: Erlenmeyer flask (100 mL) and bioreactor (7 L). The intelligent models were implemented to verify the ability to predict the emulsification index and biosurfactant concentration in smaller scale and the biosurfactant concentration and the superficial tension reduction (STR) in bigger scale, resulting in four different situations. The overall results of the predictions led to artificial neural networks as the best performing modeling tool in all four situations studied, with R2 values ranging from 0.9609 to 0.9974 and error indices close to 0. Also, four different models (Wu, Contois, Megee, and Ghose-Tyagi) were adjusted by particle swarm optimization (PSO) in order to describe the kinetics of biosurfactant production. Contois model was the only one to present R2???0.97 for all monitored variables. The findings described in this work present an adjusted model for the prediction of biosurfactant production and also state that the most adjusted kinetic model for further studies on this process is Contois model, leading to the conclusion that biomass growth is limited by a single substrate, considering only glucose.