Resumo:
The global demand for energy and the environmental impacts associated with fossil fuels have
driven the search for renewable sources, with a highlight on biodiesel. Castor oil (Ricinus
communis L.) is a promising raw material because it is non-edible and has adaptable cultivation,
although its transesterification faces common process challenges. In this context, machine
learning (ML) emerges as a tool to model and optimize this complex and non-linear process.
The general objective of this work was to develop and compare the performance of different
ML architectures for the predictive modeling and optimization of the operational parameters of
the homogeneous transesterification of castor oil, aiming to maximize biodiesel yield. Six
architectures were analyzed: Multilayer Perceptron (MLP-logsig and MLP-tansig), Radial
Basis Function Network (RBF), a hybrid model (RBF+MLP), Random Forest (RF), and
Adaptive Neuro-Fuzzy Inference System (ANFIS), using a database of 406 labeled experimental
sets from the literature. The models were evaluated using metrics such as the correlation
coefficient (R), mean square error (MSE), and root mean square error (RMSE). The MLP-tansig
model demonstrated the best predictive performance, with R > 0.98 in all phases and a test
RMSE of 3.03%. For the reverse optimization stage, a Genetic Algorithm (GA) was coupled to
the models, and the GA-RBF combination yielded the operational conditions most consistent
with the literature, despite the superior point prediction performance of the MLP-tansig model:
basic catalyst, alcohol/oil molar ratio of 19.35:1, catalyst concentration of 1.13% (w/w),
temperature of 49.91 °C, reaction time of 70.44 min, and stirring at 548.32 rpm, achieving a
predicted yield of 100% methyl esters. It is concluded that the proposed methodology is robust
and effective, integrating artificial intelligence into process engineering to optimize biodiesel
production, with potential application to other biomasses.
