@article{WANG2026182098,
title = {Reliable extrapolation of multicomponent thermophysical properties to extreme operating conditions via physics-constrained generative learning},
journal = {Chemical Engineering Journal},
volume = {548},
pages = {182098},
year = {2026},
issn = {1385-8947},
doi = {https://doi.org/10.1016/j.cej.2026.182098},
url = {https://www.sciencedirect.com/science/article/pii/S1385894726095616},
author = {Yuan Wang and Weidong Zhang and Shuaiying Yuan and Hongxin Ming and Song Li and Hui Li and Dahuan Liu},
keywords = {Multicomponent systems, Thermophysical property prediction, Wide-range extrapolation, Physics-constrained machine learning},
abstract = {Accurate macroscopic properties of multicomponent systems are a quantitative foundation for chemical process design, resource extraction, materials development, and energy conversion. Under high temperature, high pressure, and complex compositions, experimental acquisition is costly and hazardous, often leaving process design dependent on conservative estimates and costly pilot-scale validation. The central challenge is not interpolation among measured points, but reliable extrapolation from limited accessible data into experimentally demanding domains. To overcome divergent extrapolation in semi-empirical models caused by parameter non-uniqueness and out-of-distribution collapse in data-driven approaches, we propose PC-CVAE, a physics-constrained generative framework that learns thermodynamic manifolds without requiring property-specific governing equations. The framework uses the Gibbs phase rule as a physics-informed dimensional prior to guide the choice of latent dimensionality, anchors composition boundaries with lower-dimensional subsystem data, and maps operating conditions deterministically to manifold coordinates. Validated across solubility, viscosity, and thermal conductivity, PC-CVAE achieves far-range extrapolation R2 values of 0.892, 0.981, and 0.851, respectively. For solubility and viscosity, RMSE is reduced by 60% and 70% relative to classical semi-empirical baseline. Its stable transfer to thermal conductivity across a 200 °C extrapolation span indicates transferability across the three property types examined, which span fundamentally different molecular mechanisms.}
}