Physics-Guided Learning with Calibrated Uncertainty for Molecular Thermodynamics Beyond Observed Chemical Space

Scientific machine learning seeks to explore vast, high-dimensional design spaces, but experimental data are typically sparse and unevenly distributed, limiting generalization beyond observed systems. Molecular thermodynamic property prediction is a prime example. We develop a Physics-guided PC-SAFT neural NETwork, P-SANET, within a reliability-oriented framework integrating physical models, physics-guided target formulation, chemically informed out-of-distribution (OOD) evaluation and monotonic error calibration. P-SANET learns the temperature-dependent charging free energy and maps it through PC-SAFT to molecular interaction parameters. This thermodynamically structured target improves prediction accuracy for scaffold-disjoint and structurally novel molecules. A matched-size open-data benchmark further shows that chemically informative coverage, rather than nominal dataset size, is a key determinant of OOD reliability. A monotonic calibrator combining temperature, ensemble uncertainty and embedding novelty provides quantitative error estimates. P-SANET achieves competitive vapor-pressure accuracy against published models and transfers without property-specific retraining to enthalpy of vaporization, liquid density and critical properties.