Bridging Hydrocarbon Thermodynamics and Electron-Density Isosurfaces with Explainable Machine Learning.

Key thermodynamic properties of hydrocarbons are essential for process design and for reliable thermodynamic modeling. Experimental measurements are often costly or unavailable for many compounds, and prediction models are rarely further updated and optimized after the establishment of the classic thermodynamic estimation method. Here we develop compact and physically interpretable QSPR models that predict these properties from descriptors obtained by quantitative analysis of molecular surface defined on the 0.001 au electron-density isosurface. Molecular geometries and wave functions are generated using a baseline DFT protocol (B3LYP/6-31G(d,p) with harmonic frequency checks), and the robustness of the selected descriptor set is further examined across alternative functionals/basis sets and solvent-continuum settings. Using variance-inflation-factor (VIF) pruning to control multicollinearity, we combine sparse linear models (LASSO) with sure independence screening and sparsifying operator (SISSO) to obtain concise closed-form expressions (typically 1-2 composite descriptors). Across the corresponding test splits, all reported models achieve high predictive accuracy (test-set R2 generally > 0.95), while retaining clear physical meaning: density-like packing proxy (M/V), geometric size/shape measures (V, sphericity S), and, where needed, electrostatic-surface statistics (e.g., charge-balance ν, ESP skewness) jointly rationalize trends in volatility and critical behavior. The resulting formulas provide a practical bridge between first-principles electronic-structure calculations at the molecular scale and macroscopic thermodynamic properties, enabling rapid property estimation with transparent structure-property interpretation.