Ciclos sucessivos de adsorção e dessorção de Cr2+ em coluna de leito fixo: modelagem via redes neurais artificiais
Data
2022-02-04
Autores
Mello, Rebecca Ruzene 
Teixeira, Lucas Brito
Orientadores
Martins, Tiago Dias
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Com o avanço de políticas públicas nacionais e internacionais, a indústria se vê em uma conjuntura de procura por tecnologias que auxiliem no cumprimento de legislações ambientais
a favor do despejo apropriado de efluentes. O método de adsorção e dessorção por material
biológico é uma prática alternativa para o tratamento desses efluentes, com ele é possível extrair
impurezas e metais pesados do fluido contaminante e reutilizar o material filtrante após o ciclo
de biossorção. Sua vantagem sobre os métodos tradicionais são a alta eficiência, alta demanda
e, por consequência, o baixo custo de investimento. No entanto, seus modelos fenomenológicos
são difíceis de se determinar devido aos vários parâmetros que influenciam em sua atividade,
além da rigorosidade matemática a ser elaborada. Diante deste cenário, surgese a possibilidade
da aplicação dos princípios das Redes Neurais Artificiais (RNA) que consistem em um modelo
de otimização computacional com elevado poder de aprendizagem. Assim, através de uma base
de dados, as RNAs promovem a possibilidade da previsão de desempenho e oferecem uma
forma mais simples de estudo de comportamentos complexos. Com isso, este estudo focou na
aplicação dos conceitos de RNAs para a modelagem cinética de ciclos sucessivos de adsorção
e dessorção de Cr2+ em coluna de leito fixo. Os dados experimentais utilizados para o
treinamento da RNA foram extraídos da tese de doutorado de Seolatto (2008), em que se
investigou o comportamento biossorção da alga Sargassum filipendula. Um fator desafiador
neste trabalho, e enfrentado por muitos cientistas de dados, foi a quantidade de dados
disponíveis para treinamento, para tanto, os autores utilizaramse do software multiplicador de
dados TableCurve® para determinar uma curva de pontos em quantidades possíveis de
treinamento. A metodologia abordada para o processamento dos dados foi variar os parâmetros
de funções de otimização, prosseguir com duas camadas intermediárias e variar de 5 a 30
neurônios para cada camada. Esta metodologia foi adaptada a partir da experiência dos próprios
autores em seus estudos de iniciação científica. Por fim, este trabalho validou as eficiências e
limitações para ambos os fenômenos de adsorção e dessorção, em que se obteve Mean Square
Errors (MSEs) parecidos, sendo 13,3893 para a adsorção e 15,7168 para a dessorção. A partir
dos resultados, foi possível identificar que as funções de ativação Tansig e Purelin são ótimos
para este tipo de problema, representando mais de 50% das melhores sessenta RNAs.
With the advance of national and international public policies, the industry finds itself at a need of searching for technologies that help comply with environmental legislation in favor of the proper disposal of effluents. The method of adsorption and desorption by biological material is another practical method for the treatment of these effluents. With this, it is possible to remove impurities and heavy metals from the waste and to reuse the filtering material after the biosorption cycle. Its advantages over traditional methods are high efficiency, high demand and, consequently, low investment cost. However, its phenomenological models are difficult to determine due to the various parameters that influence its activity, in addition to the mathematical rigor to be elaborated. Given this scenario, the possibility of applying the principles of Artificial Neural Networks (ANN) arises, which consist of a computational optimization model with high learning power. Thus, through a database, ANNs promote the possibility of predicting performance and offer a simpler way to study complex behaviors. Thus, this study focused on the application of ANN concepts for the kinetic modeling of successive cycles of Cr2+ adsorption and desorption in a fixedbed column. The experimental data used for ANN training were extracted from the Ph.D. Dissertation of Seolatto (2008), in which the biosorption behavior of the alga Sargassum filipendula was investigated. A challenging factor in this work, and faced by many data scientists, was the amount of data available for training, so the authors used the data multiplier software TableCurve® to determine a curve of points in possible training quantities. The methodology approached for processing the data was to vary the parameters of optimization functions, proceed with two intermediate layers, and vary from 5 to 30 neurons for each layer. This methodology was adapted from the authors' own experience in their scientific initiation studies. Finally, this work validated the efficiencies and limitations for both adsorption and desorption phenomena, in which similar Mean Square Errors (MSEs) were obtained, being 13.3893 for adsorption and 15.7168 for desorption. From the results, it was possible to identify that the Tansig and Purelin activation functions are excellent for this type of problem, representing more than 50% of the best sixty ANNs.
With the advance of national and international public policies, the industry finds itself at a need of searching for technologies that help comply with environmental legislation in favor of the proper disposal of effluents. The method of adsorption and desorption by biological material is another practical method for the treatment of these effluents. With this, it is possible to remove impurities and heavy metals from the waste and to reuse the filtering material after the biosorption cycle. Its advantages over traditional methods are high efficiency, high demand and, consequently, low investment cost. However, its phenomenological models are difficult to determine due to the various parameters that influence its activity, in addition to the mathematical rigor to be elaborated. Given this scenario, the possibility of applying the principles of Artificial Neural Networks (ANN) arises, which consist of a computational optimization model with high learning power. Thus, through a database, ANNs promote the possibility of predicting performance and offer a simpler way to study complex behaviors. Thus, this study focused on the application of ANN concepts for the kinetic modeling of successive cycles of Cr2+ adsorption and desorption in a fixedbed column. The experimental data used for ANN training were extracted from the Ph.D. Dissertation of Seolatto (2008), in which the biosorption behavior of the alga Sargassum filipendula was investigated. A challenging factor in this work, and faced by many data scientists, was the amount of data available for training, so the authors used the data multiplier software TableCurve® to determine a curve of points in possible training quantities. The methodology approached for processing the data was to vary the parameters of optimization functions, proceed with two intermediate layers, and vary from 5 to 30 neurons for each layer. This methodology was adapted from the authors' own experience in their scientific initiation studies. Finally, this work validated the efficiencies and limitations for both adsorption and desorption phenomena, in which similar Mean Square Errors (MSEs) were obtained, being 13.3893 for adsorption and 15.7168 for desorption. From the results, it was possible to identify that the Tansig and Purelin activation functions are excellent for this type of problem, representing more than 50% of the best sixty ANNs.