Thermodynamic behavior of CO2–organic mixtures: Density prediction using explainable machine learning algorithms

• First application of explainable ML for CO₂–organic mixture density prediction. • Models trained on 15,428 data points covering wide thermodynamic ranges. • RBFNN achieved superior accuracy with testing MAPE below 0.7 %. • Simulations reproduced key thermodynamic behaviors, including density crossover. • SHAP analysis revealed critical pressure and molecular weight as dominant factors. Accurate knowledge of density in CO 2 –organic mixtures is essential for applications such as CO 2 storage, enhanced oil recovery, and utilization technologies. Traditional approaches, including thermodynamic equations of state and empirical correlations, often suffer from limited accuracy, narrow applicability ranges, and dependence on compound-specific fitting parameters. To overcome these limitations, this study develops advanced data-driven models capable of capturing the complex thermophysical behavior of CO 2 –organic systems across broad conditions. An extensive database of 15,428 experimental density points, across a broad range of conditions, was compiled. Three machine learning frameworks were designed and optimized: support vector machine (SVM), artificial neural network (ANN), and radial basis function neural network (RBFNN). Bayesian regularization was employed to ensure robust parameter optimization. While all models presented excellent predictions, RBFNN achieved the best performance, with a mean absolute percentage error (MAPE) of 0.66 %, and a standard deviation of 5.54 % in the testing phase. A five-fold cross-validation verified the reliability of the proposed models. Beyond numerical accuracy, the models successfully reproduced physical density variations, including pressure-induced compaction, thermal expansion, and the density crossover phenomenon at high CO₂ fractions. Outlier detection analysis confirmed that over 99 % of the data were within the applicability domain. SHapley Additive exPlanations (SHAP) analysis further revealed that organic critical properties exert the strongest influence on density, followed by operating conditions, in line with physical laws and experimental evidence. Overall, the new models provided accurate, generalizable, and physically interpretable predictive frameworks for CO 2 –organic mixture density, overcoming the limitations of conventional methods and offering a valuable tool for industrial and environmental applications.