Estimation of vaporization properties of pure substances using artificial neural networks

Abstract

Vaporization properties are important for equipment modeling and process control involving liquid-vapor equilibrium. The aim of this work was to obtain an Artificial Neural Network (ANN) to predict volume, internal energy, enthalpy, and entropy of vaporization, and the saturation pressure (or temperature) of several fluids. Two strategies were proposed: one using saturation temperature in the inputs and another using saturation pressure. Five physicochemical descriptors were used to distinguish each substance. All ANNs were trained using the Levenberg-Marquardt method. Nine outputs combination were evaluated to obtain the best model, which was determined by simulating a dataset not shown in the training/validation/test steps. The results showed that ANNs with three outputs presented higher accuracy. The best one (structure 4-40-40-3) predicted saturation pressure, internal energy and enthalpy of vaporization as outputs and presented relative errors as low as 0.02%. Finally, we showed that ANNs can be reliable for vaporization properties prediction.