Comparative evaluation of machine learning density prediction for n-Alkane + n-Alkylcyclohexane mixtures including branched isomers

Reliable density estimation of liquid mixtures is essential for fuel engineering, process design, and quality control in the petroleum and chemical industries. This work compares eight machine learning approaches — SVR with RBF kernel, Random Forest, XGBoost, MLP, KNN, Decision Tree, a deep neural network, and a convolutional neural network — for predicting the density of binary n-alkane (C₉–C₁₆)/n-alkylcyclohexane (methyl- through dodecyl-, including selected branched isomers) mixtures. A dataset of 3,970 samples with four features (temperature, alkane carbon number, solute carbon number (R), and mole fraction) was assembled from published experimental studies, with hyperparameters tuned via 5-fold grid search. XGBoost achieved the best performance (test R² = 0.9980, RMSE = 1.0634 kg/m³, MAE = 0.6302 kg/m³), followed by SVR-RBF (R² = 0.9956, RMSE = 1.5737 kg/m³), Random Forest (R² = 0.9894, RMSE = 2.4307 kg/m³), and MLP (R² = 0.9875, RMSE = 2.6431 kg/m³). 10-fold cross-validation confirmed XGBoost's robustness (mean R² = 0.9969 ± 0.0011), with a bootstrap 95% confidence interval of [0.9968, 0.9989] for test R². Permutation importance identified mole fraction (0.8598) and temperature (0.6018) as the dominant predictors, with alkane carbon number contributing least (0.2252). Leave-one-alkane-out validation demonstrated robust generalization (mean R² = 0.9423), except for n-hexadecane (R² = 0.7682), while leave-one-R-out validation showed weakest performance for R=1 (R² = 0.6345). These findings can facilitate the development of accurate surrogate fuel formulations and provide a reliable data-driven framework for density estimation within the validated chemical space (n-alkanes C₉–C₁₆, n-alkylcyclohexanes with R = 1–12 and selected branched isomers, T = 288.15–373.15 K, and 0.1 MPa).