Machine Learning na predição do acúmulo de carboidratos pela microalga Chlorella vulgaris em cultivo contínuo

Abstract

In this study, Machine Learning (ML) technique was applied to predict carbohydrate production by the microalgae Chlorella vulgaris in a continuous cultivation, using nutritional (nitrogen concentration, 500-75 mg L-1, and phosphorus, 200-50 mg L-1, in medium), environmental (light intensity between 150-450 µmol m-2 s-1 and optical density between 0.5-20) and operational (residence time between 2.9-4.4 days and normalized time between 0-1) variables as inputs for the predictive models on biomass and carbohydrate productivity and the % of carbohydrate accumulation in biomass. The analysis of Pearson's coefficient allowed to identify significant correlations between the independent and dependent variables, highlighting the influence of nutrient concentration and light on carbohydrate production. A high correlation was observed between nitrogen and phosphorus, which can compromise the quality of the regressions due to collinearity; as an alternative, the N/P ratio was used to circumvent this problem. The histogram of the variables showed patterns and trends in the data, impacting the predictions. Different regression techniques were implemented and compared, including linear models (Multivariate Linear Regression, Ridge and LASSO) and non-linear models (Random Forest, Artificial Neural Networks and Support Vector Regression – SVR). The non-linear models showed better performance in predicting all output variables, especially Random Forest and Neural Networks, which captured complex relationships between the variables. The best performance was obtained with Random Forest, reaching R² of 0.9347 and RMSE of 0.2556 for carbohydrate productivity, and R² of 0.8962 and RMSE of 0.3222 for biomass productivity. The optimization of the models was performed through grid search, allowing the definition of the best hyperparameters and significantly improving the accuracy of the predictions. Furthermore, cross-validation was used to avoid overfitting and ensure generalizability of the results, making the models more robust and reliable. In addition to the statistical analysis of the models, a graphical evaluation of the residuals was performed to verify the adequacy of the predictions. It was observed that, despite the good numerical performance of the nonlinear models, the residuals indicated a slight centralized bias, suggesting that improvements can be achieved with the inclusion of new input variables or adjustments in the data preprocessing.