Solubility prediction of 21 azo dyes in supercritical carbon dioxide using wavelet neural network

The solubility of 21 azo dyes in supercritical carbon dioxide was related to the six descriptors over a wide range of pressures (100–355 bar) and temperatures (308–413 K). The wavelet neural network (WNN) model was constructed with six descriptors as an input layer, eight neurons as a hidden layer and a neuron as an output layer. The descriptors consisted of temperature, pressure, LUMO energy, polarizability, volume of the molecule and number of unsaturated bonds and they were selected based on stepwise feature selection from different descriptors using multiple linear regression (MLR) method. The WNN architecture and its parameters were optimized simultaneously. The data were randomly divided into the training, prediction and validation sets. The RMSE and mean absolute errors in WNN model were 0.220 and 0.158 for prediction set and 0.156 and 0.114 for validation set. In addition, the prediction ability of the model was also evaluated for five azo dyes, the molecules and data of which were not in any previous data sets.The performance of the WNN model was also compared with artificial neural network (ANN) and MLR models.