Uso de redes neurais artificiais na modelagem cinética de extração supercrítica e comparação com modelos fenomenológicos
Data
2020-10-15
Autores
Irizawa, Ian de Alencar 
Orientadores
Veggi, Priscilla Carvalho
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
A artemisinina é o composto majoritário sintetizado a partir da Artemísia annua L.
de grande interesse para as indústrias farmacêuticas, cosméticas e/ou alimentícia. Sua principal
característica é sua propriedade antimalárica, descoberta esta que rendeu um Prêmio Nobel. Em
se tratando do emprego da extração com fluidos supercríticos na obtenção destes compostos,
inúmeras pesquisas vêm empregando modelos matemáticos para descrever o comportamento
cinético desse processo, que se destaca por ser sustentável, uma green technology. Neste
sentido, este trabalho teve como objetivo o desenvolvimento de uma rede neural artificial para
modelar a cinética de extração supercrítica de artemisinina. Foram utilizados oito experimentos
com diferentes condições operacionais como base de dados. Para o desenvolvimento da rede,
foram traçadas duas estratégias a fim de se obter uma curva cinética com a massa de extrato de
artemisinina em função do tempo. Na primeira estratégia do treinamento da rede, utilizaram-se
como variáveis de entrada a pressão de operação, temperatura do solvente, vazão de solvente e
massa de extrato nos tempos t e t-1. A variável de saída foi a massa de extrato no tempo t+1.
Na segunda estratégia se utilizou como variáveis de entrada a pressão, temperatura, vazão do
solvente e o tempo, tendo como variável de saída a massa de extrato no tempo t. Foram testadas
diversas configurações e avaliou-se o erro médio percentual da simulação, bem como a
correlação de Pearson a seleção da melhor rede. A melhor RNA foi obtida a partir da segunda
estratégia, com uma estrutura contendo 7 neurônios na primeira camada intermediária e 1
neurônio na segunda (estrutura 4-7-1-1), com funções de ativação Purelin-Tansig-Tansig. Essa
rede foi capaz de descrever e predizer de maneira precisa a cinética da extração supercrítica da
artemisinina com uma alta correlação de Pearson de 0,997 e um baixo erro médio na simulação
de, aproximadamente, 5%. Além disso, ao compararmos o erro médio quadrático obtido pelo
melhor dos modelos fenomenológicos estudados, esta obteve um erro de 2,358. 10-1
, valor este
extremamente inferior ao obtido pela RNA 4-7-1-1, que obteve um erro excepcionalmente
baixo, 2,619.10-3
. Com isso, pôde-se comprovar a eficiência da rede neural desenvolvida e sua
excelente capacidade de generalização do processo.
Artemisinin is the major compound synthesized from Artemísia annua L. of great interest for the pharmaceutical, cosmetic and / or food industries. Its main characteristic is its antimalarial property, a discovery that won a Nobel Prize. Regarding the use of extraction with supercritical fluids to obtain these compounds, numerous researches have been using mathematical models to describe the kinetic behavior of this process, which stands out for being sustainable, a green technology. In this sense, this work aimed to develop an artificial neural network to model the kinetics of supercritical extraction of artemisinin. Eight experiments with different operational conditions were used as a database. For the development of the network, two strategies were devised in order to obtain a kinetic curve with the mass of artemisinin extract as a function of time. In the first network training strategy, operating pressure, solvent temperature, solvent flow rate and extract mass at times t and t-1 were used as input variables. The output variable was the extract mass at time t + 1. In the second strategy, pressure, temperature, solvent flow and time were used as input variables, with the extract mass at time t as the output variable. Several configurations were tested and the average percentage error of the simulation was evaluated, as well as Pearson's correlation to the selection of the best network. The best RNA was obtained from the second strategy, with a structure containing 7 neurons in the first intermediate layer and 1 neuron in the second (structure 4-7-1-1), with Purelin-Tansig-Tansig activation functions. This network was able to accurately describe and predict the kinetics of the supercritical extraction of artemisinin with a high Pearson correlation of 0.997 and a low average error in the simulation of approximately 5%. In addition, when comparing the mean square error obtained by the best of the studied phenomenological models, it obtained an error of 2.358. 10-1 , a value that is extremely lower than that obtained by RNA 4-7-1-1, which obtained an exceptionally low error, 2,619.10-3 . With that, it was possible to prove the efficiency of the developed neural network and its excellent ability to generalize the process.
Artemisinin is the major compound synthesized from Artemísia annua L. of great interest for the pharmaceutical, cosmetic and / or food industries. Its main characteristic is its antimalarial property, a discovery that won a Nobel Prize. Regarding the use of extraction with supercritical fluids to obtain these compounds, numerous researches have been using mathematical models to describe the kinetic behavior of this process, which stands out for being sustainable, a green technology. In this sense, this work aimed to develop an artificial neural network to model the kinetics of supercritical extraction of artemisinin. Eight experiments with different operational conditions were used as a database. For the development of the network, two strategies were devised in order to obtain a kinetic curve with the mass of artemisinin extract as a function of time. In the first network training strategy, operating pressure, solvent temperature, solvent flow rate and extract mass at times t and t-1 were used as input variables. The output variable was the extract mass at time t + 1. In the second strategy, pressure, temperature, solvent flow and time were used as input variables, with the extract mass at time t as the output variable. Several configurations were tested and the average percentage error of the simulation was evaluated, as well as Pearson's correlation to the selection of the best network. The best RNA was obtained from the second strategy, with a structure containing 7 neurons in the first intermediate layer and 1 neuron in the second (structure 4-7-1-1), with Purelin-Tansig-Tansig activation functions. This network was able to accurately describe and predict the kinetics of the supercritical extraction of artemisinin with a high Pearson correlation of 0.997 and a low average error in the simulation of approximately 5%. In addition, when comparing the mean square error obtained by the best of the studied phenomenological models, it obtained an error of 2.358. 10-1 , a value that is extremely lower than that obtained by RNA 4-7-1-1, which obtained an exceptionally low error, 2,619.10-3 . With that, it was possible to prove the efficiency of the developed neural network and its excellent ability to generalize the process.