Aprendizado de máquina para a determinação das composições de sistemas em equilíbrio químico e de fases: uma comparação com um método de minimização da energia livre de Gibbs

Abstract

Chemical and physical equilibrium are crucial in the analysis of fluid phases in various separation processes; however, the calculation presents certain complexities and is computationally costly. Methods for solving this problem are based on stoichiometric and nonstoichiometric calculations and on Gibbs free energy minimization. The Gibbs free energy minimization (min G) method stands out by simultaneously solving the problem and providing phase composition solutions in a less computationally intensive way. For finding the global minimum point, min G requires the function to be convex. Thus, recently, the geometric programming (GP) method has been applied to the min G problem, ensuring convergence regardless of the function's convexity. In this context, machine learning offers a promising alternative, with artificial neural networks (ANNs) accurately and quickly predicting equilibrium values. Therefore, this work aims to analyze the performance of two types of neural networks, the multilayer perceptron (MLP) and the radial basis function network (RBF), for the simultaneous solution of chemical and phase equilibrium, applied to two typical chemical engineering problems. The input data for training the networks were generated through the min G baseline solution using the GP algorithm. Finally, the results indicated that ANNs are effective tools for this type of calculation, with the MLP standing out for its performance in both problems addressed, showing a lower mean squared error (MSE) and higher coefficient of determination (R2). The RBF ANN showed inferior results in both cases due to its single-layer limitation and difficulty handling non-linearly dependent data.