Applying artificial neural networks for vapor-liquid equilibrium prediction

• Main Author: Nguyen, Viet D.
• Format: text
• Language: English
• Published: Animo Repository 2006
• Subjects: Vapor-liquid equilibrium; Artificial neural networks; Chemical Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/3543; https://animorepository.dlsu.edu.ph/context/etd_masteral/article/10381/viewcontent/CDTG004353_P.pdf
• Institution: De La Salle University

Most solvents used in the semiconductor industry are toxic and costly. Thus, the solvents should be recovered for re-use in these processes by distillation methods, and vapor liquid equilibrium (VLE) data are necessary for the design and operation of distillation columns. The VLE data can be estimated using thermodynamic models based on calculation activity coefficients. However, while thermodynamic models can apply well to ideal systems, it is difficult to apply for non-ideal solvent systems, especially for multi-component systems with the effect of salt. Artificial neural networks (ANNs), on the other hand, are purely numerical methods (no theoretical interpretation possible) but are still useful for design applications. ANNs can overcome the limitations of the traditional methods by extracting the desired information directly from the experimental data. ANNs have become established method with chemical engineering. In this work, ANNs were applied to predict and estimate VLE data for binary and ternary systems without and with salts (five binary systems without and with salt two ternary systems saturated with salt). The database was taken from literature and split into two subsets: training and validating data. For the VLE of ethanol-toluene-sodium acetate in the pressure range of 77- 757 mmHg, the mean absolute deviations (MADs) in vapor mole fraction and bubble points for the whole database were 0.0048 and 0.35oC, respectively. In the case of ternary systems, the components are isopropyl alcohol (IPA), 1-propanol and ethanol. These solvents and water are used extensively in various stages of semiconductor product washing and cleaning. The MADs for entire database were 0.0169, 0.0166, 0.0147 and 0.72oC in vapor mole fraction of components (y1, y2, y3) and bubble points, respectively. The VLE data predicted by thermodynamic models (Wilson and Tan-Wilson) and ANNs were compared with the published experimental data. The ANNs showed better agreement with published experimental data than the thermodynamic models.