Answers are grounded in this paper and cite their sources.
Study benchmarks AI models against classical thermodynamic models for predicting phase behavior in binary systems, suggesting AI's potential in thermodynamics.
Artificial intelligence offers fast alternatives to traditional mechanistic thermodynamic modeling, yet its ability to replace established frameworks remains uncertain. This work benchmarks emerging deep-learning architectures (SPT-NRTL and GNN-based models) against UNIFAC(Do) and COSMO-SAC across a chemically diverse set of 95 binary VLE systems. We assess whether AI-predicted infinite-dilution activity coefficients (IDACs) can serve as a robust thermodynamic property for data-efficient general phase-equilibrium calculations. Under isothermal conditions, IDAC-based reconstructions using GNN-derived activity coefficients yielded VLE predictions with an MRE of approximately 17%–22%, while SPT-NRTL (10.5%) captured even complex azeotropes, performing competitively with COSMO-SAC (12.6%) and UNIFAC(Do) (8.6%). GNNs showed systematic deviations for strongly polar mixtures. Under isobaric conditions, SPT-NRTL showed better agreement than models based on IDACs at 298 K.
No takes yet. Share an insight, caveat, or question.
Schneider et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: