Perfil cinético da adsorção da mistura Ni2+-Cr3+ via redes neurais artificiais: avaliação de diferentes algoritmos de otimização
Nenhuma Miniatura disponível
Data
2021-02-19
Autores
Choji, Thamyres Tetsue [UNIFESP]
Orientadores
Martins, Tiago Dias [UNIFESP]
Tipo
Trabalho de conclusão de curso de graduação
Título da Revista
ISSN da Revista
Título de Volume
Resumo
A preocupação mundial com o tratamento de resíduos produzidos pela indústria é uma realidade. A adsorção é um dos processos mais utilizados especialmente em etapas de refino, em que as concentrações dos componentes em questão são baixas e processos convencionais não são eficientes. A definição de um modelo cinético binário é complexa e envolve diversas variáveis independentes. Sendo assim as Redes Neurais Artificiais podem ser uma alternativa viável, já que não exigem grande conhecimento matemático acerca do fenômeno em questão e possuem a versatilidade necessária para se adaptar a diferentes situações. Neste projeto, o objetivo foi obter e comparar modelos neurais que descrevam o perfil cinético de adsorção da mistura Ni2+-Cr3+ pela alga marinha Sargassum filipendula no intuito de suprir as dificuldades encontradas na formulação dos modelos fenomenológicos. Um estudo sistemático foi realizado para se determinar os números de camadas ocultas e de neurônios, tal que sejam suficientes para descrever o fenômeno em questão. Para isso, foram comparados o desempenho de diversos algoritmos de otimização em busca da melhor configuração. Os resultados mostraram que o método de otimização Levenberg-Marquardt com Regularização bayesiana (trainbr) com funções de transferências tangente hiperbólica para estrutura 4-10-10-2 foi que melhor se adaptou aos dados de adsorção binária para os metais em questão, apresentando função objetivo mais satisfatória (6,1.10-10) dentre os modelos testados. Além disso, foi capaz de simular o perfil cinético de adsorção utilizando na entrada da rede apenas os dois primeiros pontos da curva e utilizando os pontos gerados na retroalimentação. Portanto, demonstrou o poder de generalização das RNAs em um problema de adsorção binária.
The worldwide concern with the treatment of waste produced by the industry is a reality. Adsorption is one of the most used processes, especially in refining steps, in which the concentrations of the components in question are low and conventional processes are not efficient. The definition of a binary kinetic model is complex and involves several independent variables. Thus, Artificial Neural Networks can be a viable alternative, since they do not require great mathematical knowledge about the phenomenon in question and have the necessary versatility to adapt to different situations. In this project, the objective was to obtain and compare neural models that describe the kinetic profile of adsorption of the Ni2 + -Cr3 + mixture by the seaweed Sargassum filipendula in order to overcome the difficulties found in the formulation of phenomenological models. A systematic study was carried out to determine the numbers of hidden layers and neurons, such that they are sufficient to describe the phenomenon in question. For that, the performance of several optimization algorithms were compared in search of the best configuration. The results showed that the Levenberg-Marquardt optimization method with Bayesian Regularization (trainbr) with hyperbolic tangent transfer functions for structure 4-10-10-2 was the one that best adapted to the binary adsorption data for the metals in question, presenting a function most satisfactory objective (6.1, 10-10) among the tested models. In addition, it was able to simulate the adsorption kinetic profile using only the first two points of the curve at the entrance of the network and using the points generated in the feedback. Therefore, it demonstrated the generalization power of RNAs in a binary adsorption problem.
The worldwide concern with the treatment of waste produced by the industry is a reality. Adsorption is one of the most used processes, especially in refining steps, in which the concentrations of the components in question are low and conventional processes are not efficient. The definition of a binary kinetic model is complex and involves several independent variables. Thus, Artificial Neural Networks can be a viable alternative, since they do not require great mathematical knowledge about the phenomenon in question and have the necessary versatility to adapt to different situations. In this project, the objective was to obtain and compare neural models that describe the kinetic profile of adsorption of the Ni2 + -Cr3 + mixture by the seaweed Sargassum filipendula in order to overcome the difficulties found in the formulation of phenomenological models. A systematic study was carried out to determine the numbers of hidden layers and neurons, such that they are sufficient to describe the phenomenon in question. For that, the performance of several optimization algorithms were compared in search of the best configuration. The results showed that the Levenberg-Marquardt optimization method with Bayesian Regularization (trainbr) with hyperbolic tangent transfer functions for structure 4-10-10-2 was the one that best adapted to the binary adsorption data for the metals in question, presenting a function most satisfactory objective (6.1, 10-10) among the tested models. In addition, it was able to simulate the adsorption kinetic profile using only the first two points of the curve at the entrance of the network and using the points generated in the feedback. Therefore, it demonstrated the generalization power of RNAs in a binary adsorption problem.