A neural networks method to predict activity coefficients for binary systems based on molecular functional group contribution.

摘要

Artificial neural network (ANN) techniques and functional group contributions were used to develop an algorithm to predict chemical activity coefficients. The ANN algorithm was trained using experimental data for more than 900 binary systems obtained from DECHEMA, a phase-equilibrium database. All experimental data binary systems used in this study are isothermal. The prediction scheme is based on the fact that the atoms in a chemical compound can be grouped in a functional group with its own physical and chemical properties. Thus, almost any chemical compound can be built by combining the right number of functional groups. The functional group interactions among the components in a mixture are estimated and the combination of functional group interactions provides the intermolecular relationship among the components of a mixture and consequently the activity coefficients can be predicted. The intramolecular interactions were not considered in this study. The four-suffix Margules equation was used as the base thermodynamic model to calculate the activity coefficients. The Margules equation is good for modeling enthalpic contributions to the activity coefficient but is not good for modeling entropic contributions to the activity coefficient. The design of functional groups based on quantum mechanics was adopted to develop a method for predicting activity coefficients. ANN techniques are especially useful for modeling a highly nonlinear interaction among the functional groups and the corresponding activity coefficient. One of the major contributions of this research is to propose a method to identify the initial point and the structure of an ANN. The minimum mean squared prediction error criterion was implemented to determine both a suitable initial point and the structure of the ANN. A random search method was used to determine the optimal initial point and the Levenberg-Marquardt algorithm was used to train the ANN to generate a sample of prediction values and the trim mean based on 20% data elimination was selected as the best representation of the ensemble prediction of the Margules equation parameters. The algorithm was validated with nineteen vapor-liquid equilibrium systems and results show that the ANN provides a relative improvement over the UNIFAC method. The scope of this study is limited to some chemical compound families (i.e. alcohols, phenols, aldehydes, ketones and ethers), it is required to include more experimental data to cover additional chemical compound families such as carboxylic acids, anhydrides, esters, aliphatic hydrocarbons and halogens.