The need for sustainable alternatives in refrigeration has grown as Europe enforces mandates on avoiding high global warming potential (GWP) refrigerants. CO₂-based refrigerants have emerged as a promising choice in response, distinguished by its low GWP and reduced flammability, compared to formulated hydrofluoroolefins, thus offering a safer and sustainable solution in the context of next generation drop-in refrigerants. This study presents a machine-learning-based methodology to estimate the saturation properties of CO2-based mixtures, allowing for the precise tuning of molecular-based models like the polar soft-SAFT, used for technical evaluation, without relying on experimental data, often unavailable for such systems. The approach departs from the thermodynamic characterization of several pure-components, including novel fluorine-based refrigerants. The parametrization allows an excellent description of the vapor pressure, saturated densities, and latent heat. Next, a constant, temperature-independent binary parameter is used to estimate the solubility profiles of CO2-derived mixtures in selected refrigerants. The model effectively captures azeotropic and zeotropic behaviors, demonstrating its strength in fine-tuning solubility with minimal corrections. Subsequently, data from the molecular characterization via polar soft-SAFT is used as output targets to train a machine learning algorithm based on artificial neural networks, enabling the prediction of mixture saturation properties out of the training dataset's scope. Using COSMO σ-profiles, the developed ANN demonstrates high efficiency in predicting saturation bubble and dew temperatures, achieving R² > 0.9999, RMSE< 0.0959, AARD < 0.0220 %, and NMAD of 0.00044. Statistical analysis confirms minimal mean deviations, with outliers limited to 2.63 % for bubble and 2.44% for dew phase predictions, respectively.