Correlation Models of Three Machine Learning Types for Carbon Dioxide/Toluene and Carbon Dioxide/Methanol Binary Systems

Abstract

Existing density correlation models for CO₂/organic solvent homogeneous mixture fluids face limitations in temperature and composition applicability. This study applies three machine learning methods─support vector machine regression, artificial neural networks (ANNs), and genetic programming─to develop a density correlation model effective across wide ranges of composition, pressure, and temperature. Training data were gathered by measuring the densities of CO₂/toluene (Tol) and CO₂/methanol (MeOH) binary systems using a high-pressure oscillating density meter. The measurements were conducted at a temperature range of 313–353 K, a CO₂ mole-fraction range of 0–80 mol %, and at pressures up to 20 MPa. Initially, CO₂/Tol density data were used to optimize each machine learning model’s hyperparameters. These optimized parameters enabled predictions, allowing comparison of the accuracy and interpolation performance of each method. Results showed that an ANN model, using a softsign transfer function and six neurons in the hidden layer, provided optimal accuracy and predictive range. The root-mean-square errors at this time were 4.26 and 4.71 kg m^–3 for training and validation, respectively. Machine learning with CO₂/MeOH data similarly produced reliable density predictions across broad conditions, expanding the model’s practical use in various systems.