(37a) Estimation of Activity Coefficients in Binary Systems Using Neural Networks

Artificial neural networks (ANN) techniques and a group-contribution approach were used to develop an algorithm to predict activity coefficients in binary solutions. The ANN algorithm was trained using experimental data obtained from DECHEMA, a vapor-liquid-equilibrium (VLE) database. The functional group selection was based on quantum mechanics. ANN techniques were used because they are especially useful to model highly nonlinear interactions among the functional groups and hence to predict activity coefficients. One of the major contributions of this research, from a computational point of view, is the proposed method to identify the initial point and the structure of the ANN. The minimum variance criterion was introduced 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. The Levenberg-Marquardt algorithm was applied to train the ANN to generate a sample of predicted values. The trim mean based on 20% data elimination was selected as the best representation of the ensemble prediction of the activity coefficient. The major contribution of this work, from an application point of view, is the new approach to estimate activity coefficients in the absence of VLE experimental data. The algorithm was validated with nineteen VLE systems not used to train the ANN and the results show that the ANN provides a relative improvement over the UNIFAC method.