A machine learning-based platform for dye solubility in supercritical carbon dioxide: Classification optimization and predictive analysis

Abstract

This study addresses the critical need for the accurate prediction of dye solubility in supercritical carbon dioxide (scCO₂) to optimize the dyeing process. An open-access dye solubility database was developed that includes 2991 experimental data points for 90 dyes. Machine learning models, namely the multilayer perceptron (MLP), wavelet neural network (WNN), recurrent neural network (RNN), and long short-term memory (LSTM) models, were trained to predict dye solubility. Initial mixed-data training showed limited performance, while classification based on molecular structures by separating anthraquinone dyes, azo dyes and other dyes into distinct datasets, greatly improved the prediction accuracy. For example, WNN achieved an R² value of 0.998 for anthraquinone dyes, while MLP and RNN showed better performance for azo and other dye categories. This classification strategy not only improves model performance, but also verifies the critical role of molecular structure in predicting dye solubility, offering a reliable approach for the development of accurate prediction models.