Aplicação da técnica de redes neurais artificias na modelagem de processos oxidativos avançados
Data
2022-12-14
Autores
Silva, Evelin Karolina Gomes 
Araújo, Fauzer Henrique Lopes de
Orientadores
Moraes, José Ermírio Ferreira de
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Uma grande quantidade de produtos farmacêuticos é consumida pela sociedade devido a sua eficiência no tratamento e prevenção de doenças. Porém, uma parcela dos fármacos consumidos não é metabolizada pelos seres vivos, sendo excretada nas redes de esgoto e no meio ambiente. Além disso, no caso particular dos antibióticos, como a amoxicilina, a sua presença no meio ambiente pode promover o desenvolvimento de bactérias multirresistentes. Nesse contexto, observa-se a necessidade do estudo de novos processos adequados para a degradação dessas substâncias, como é o caso dos processos oxidativos avançados (POA). Dentre os POA, pode-se destacar o processo foto-Fenton, que envolve o uso de sais ferrosos e do peróxido de hidrogênio na presença de uma fonte de radiação ultravioleta-visível (UV-Vis), promovendo a geração de radicais livres, especialmente os radicais hidroxila, que são os responsáveis pela degradação dos poluentes alvo. Dessa forma, a modelagem fenomenológica dos POA pode ser bastante complexa, pois requer as informações das cinéticas das várias reações químicas envolvidas e, além disso, o conhecimento do modelo do campo de radiação aplicado. Nesse contexto, o uso de técnicas empíricas, como as redes neurais artificiais (RNA) se apresenta como uma interessante alternativa. As RNA são capazes de buscar um padrão entre dados de saída e de entrada. Este trabalho procurou otimizar uma RNA capaz de modelar o processo foto-Fenton aplicado ao processo de degradação da amoxicilina, utilizando-se a extensão Deep Learning Toolbox disponível no software Matlab. A modelagem ocorreu a partir da otimização dos parâmetros da rede, como: a escolha entre os algoritmos de treinamento Regularização Bayesiana e Levenberg-Marquardt e o número de neurônios na camada oculta. A RNA otimizada possui 4 neurônios na camada oculta e foi criada utilizando o algoritmo Regularização Bayesiana com o auxílio de modelos empíricos polinomiais para a geração de um maior número de dados. A rede apresentou um bom grau de generalização e sem problemas de sobreajuste, com R² igual a 0,99836. Com o uso do modelo, verificou-se que o sistema apresentou maior sensibilidade à [Fe2+] em comparação à taxa molar de alimentação de peróxido de hidrogênio para o domínio experimental adotado.
A large amount of pharmaceutical products is consumed by society due to their efficiency in the treatment and prevention of diseases. However, a portion of the drugs consumed is not metabolized by living beings, being excreted in sewage networks and in the environment. Furthermore, in the case of antibiotics, such as amoxicillin, their presence in the environment can promote the development of multidrug-resistant bacteria. In this context, there is a need to study new processes suitable for the degradation of these substances, as is the case of advanced oxidative processes (AOP). Among the AOP, the photo-Fenton process can be highlighted, which involves the use of ferrous salts and hydrogen peroxide in the presence of a source of ultraviolet-visible radiation (UV-Vis), promoting the generation of free radicals, especially hydroxyl radicals, which are responsible for the degradation of target pollutants. Thus, the phenomenological modeling of AOP can be quite complex, as it requires information on the kinetics of the various chemical reactions involved and, in addition, knowledge of the applied radiation field model. In this context, the use of empirical techniques, such as artificial neural networks (ANN) are an interesting alternative. ANN can look for a pattern between output and input data. This work sought to optimize an ANN capable of modeling the photo-Fenton process applied to the amoxicillin degradation process, using the Deep Learning Toolbox extension available in Matlab software. Modeling was based on the optimization of network parameters, such as: the choice between Bayesian Regularization and Levenberg-Marquardt training algorithms and the number of neurons in the hidden layer. The optimized ANN has 4 neurons in the hidden layer and was created using the Bayesian regularization algorithm with the aid of polynomial empirical models to generate a greater amount of data. The network showed a good degree of generalization and no overfitting problems, with R² equal to 0.99836. With the use of the model, it was verified that the system presented greater sensitivity to [Fe2+] compared to the molar rate of hydrogen peroxide feed for the adopted experimental domain.
A large amount of pharmaceutical products is consumed by society due to their efficiency in the treatment and prevention of diseases. However, a portion of the drugs consumed is not metabolized by living beings, being excreted in sewage networks and in the environment. Furthermore, in the case of antibiotics, such as amoxicillin, their presence in the environment can promote the development of multidrug-resistant bacteria. In this context, there is a need to study new processes suitable for the degradation of these substances, as is the case of advanced oxidative processes (AOP). Among the AOP, the photo-Fenton process can be highlighted, which involves the use of ferrous salts and hydrogen peroxide in the presence of a source of ultraviolet-visible radiation (UV-Vis), promoting the generation of free radicals, especially hydroxyl radicals, which are responsible for the degradation of target pollutants. Thus, the phenomenological modeling of AOP can be quite complex, as it requires information on the kinetics of the various chemical reactions involved and, in addition, knowledge of the applied radiation field model. In this context, the use of empirical techniques, such as artificial neural networks (ANN) are an interesting alternative. ANN can look for a pattern between output and input data. This work sought to optimize an ANN capable of modeling the photo-Fenton process applied to the amoxicillin degradation process, using the Deep Learning Toolbox extension available in Matlab software. Modeling was based on the optimization of network parameters, such as: the choice between Bayesian Regularization and Levenberg-Marquardt training algorithms and the number of neurons in the hidden layer. The optimized ANN has 4 neurons in the hidden layer and was created using the Bayesian regularization algorithm with the aid of polynomial empirical models to generate a greater amount of data. The network showed a good degree of generalization and no overfitting problems, with R² equal to 0.99836. With the use of the model, it was verified that the system presented greater sensitivity to [Fe2+] compared to the molar rate of hydrogen peroxide feed for the adopted experimental domain.