Aplicação de aprendizado de máquina na predição do rendimento da fermentação de alginase

Abstract

Artificial intelligence techniques are important tools to model and optimize the variables that influence fermentation processes. The performance of fermentation processes is affected by many factors, including temperature, moisture content, agitation, inoculum level, carbon and nitrogen sources, etc. To achieve the best performance of fermentation processes, several process optimization strategies have been developed. This work aims to apply machine learning techniques to predict the yield of semi-solid fermentation using filamentous fungus for alginase production using brown macroalgae as substrate. The macroalgae chosen was the genus Sargassum, and the fungal strain was Cunninghamella echinulata. In total, about 115 independent experimental data were obtained from the variation of factors: cultivation time (1-7 days), humidity (65-85% m/m), inoculum concentration (2,106-1,107 spores/gbiomass), % of sodium alginate (inducer, 0-33.33% m/m) and pH of the nutrient solution (2.5-8.5), being the input variables and only the alginase enzymatic activity as the output variable. In this work, for the execution of the models, simulation and optimization, the programming language in Python was used using SciPylibrary. The identification of a non-linear relationship between fermentation, factors and desired objectives was carried out, first, using the learning capabilities of Artificial Neural Networks (ANN) and then by Support Vector Macchine (SVM), evaluating the predictive capacity through the coefficient of determination (R2). For the RNA, with 15 neurons and using a logistic activation function, it was possible to fix a model with R2 = 0.877, while for the SVM with a polynomial Kernel function, a model with R2 = 0.821 was obtained, which, although a little lower, presented similar regression. In both models, the variables that exerted influence were the cultivation time, humidity, % of sodium alginate (inducer) and pH of the nutrient solution, with the inoculum concentration in the studied interval not having a significant influence. In both methods, the maximum R2 found was considered satisfactory given the non-linearity of the fungal semi-solid fermentation for enzymatic excretion.