Answers are grounded in this paper and cite their sources.
Infinite-dilution activity coefficients (IDACs) provide a thermodynamic basis for comparing solvents, but the reliability of learned predictions depends on the chemistry available during training. We evaluated descriptor models and a Gibbs–Helmholtz graph neural network using 11,832 development observations spanning 541 molecular identities and a reserved literature panel of 2809 binary systems. For ExtraTrees with interaction descriptors (ET-I), mean absolute error in ln γ∞ increased from 0.112 under random splitting to 0.238 for unseen pairs, 0.265 for unseen solutes, and 0.327 for unseen solvents. When both identities were unseen, mean error was 0.673 ± 0.480 across five partitions (median 0.434); only the highest-error partition contained water-solvent test records. The graph model reduced mean error to 0.206 for pair holdout and 0.593 for double holdout, with 70% of its summed double-holdout improvement concentrated in that water-containing partition. Repeated training preserved the lower mean errors relative to ET-I but changed the ordering of models in individual partitions. Sensitivity analyses showed that large replicate disagreements did not account for the water error. Leading physical models were more accurate than ET-I on matched literature systems. Solvent-switch direction agreement increased with predicted contrast magnitude in retrospective comparisons, although candidate ranking uncertainty was not calibrated. Chemical support, partition variability, and contrast-level performance therefore determine how these predictions can inform solvent comparisons beyond familiar systems.
No takes yet. Share an insight, caveat, or question.
Shi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: