Quercetin extraction from Rosa damascena Mill via supercritical CO2: Neural network and adaptive neuro fuzzy interface system modeling and response surface optimization

• Published in: The Journal of supercritical fluids Vol. 112; pp. 57 - 66
• Main Authors: Ghoreishi, S.M., Hedayati, Ali, Mousavi, S.O.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2016
• Subjects: Adaptive neuro-fuzzy inference system; Artificial neural network; Quercetin; Response surface methodology; Rosa damascena Mill; Supercritical extraction; Supercritical extraction; Artificial neural network; Response surface methodology; Adaptive neuro-fuzzy inference system; Rosa damascena Mill; Quercetin
• ISSN: 0896-8446; 1872-8162
• DOI: 10.1016/j.supflu.2016.02.006

•SC-CO2 was used for Quercetin extraction from Rosa damascena Mill.•Response surface method was applied to optimize effective operating conditions.•Quercetin supercritical extraction recovery was predicted via ANN and ANFIS models.•Developed models were validated by experimental Quercetin SFE data. In this study, the extraction of quercetin from Rosa damascena Mill was carried out by modified supercritical CO2 with ethanol an entrainer and Soxhlet extraction. Design of experiment was carried out with response surface methodology (RSM) using Mini Tab software 17. The operating temperature (35–55°C), pressure (10–30MPa), dynamic extraction time (40–120min) and CO2 flow rate (0.3–1.5ml/min) were considered as the range of operating variables. Response surface analysis verified that R2 and modified R2 of the model were 93.1% and 87.1%, respectively. Optimal operating conditions was predicted using RSM modeling to be the pressure of 25.5MPa, temperature of 46.3°C, CO2 flow rate of 0.7ml/min and dynamic extraction time of 120min in which the maximum recovery of 32.0% was obtained. Moreover, the recovery of extraction was modeled by adaptive neuro-fuzzy inference system (ANFIS) and artificial neural network (ANN). Levenberg–Marquardt backpropagation training function with six neurons in hidden layer was found to be the most suitable network and the coefficient of determination (R2) was 99.5%. Gaussian curve built-in membership function using 2 membership functions to each input was obtained to be optimum ANFIS architecture with mean square error (MSE) of 0.19, 0.69 and 0.49 for training, testing and checking data, respectively.