What's the best predictive model for ionic liquid screening? COSMO-SAC, COSMO-RS and GNN

Abstract

Screening ionic liquids (ILs) is a ranking problem before it is a prediction problem: among millions of cation–anion combinations, what matters is which candidates reach the laboratory, not the value assigned to each. Models are nevertheless compared almost exclusively on absolute error. We benchmark six physics-based variants and one data-driven model against ~55,000 experimental infinite-dilution activity coefficients (γ∞) in ILs, mostly imidazolium-based with hydrocarbon solutes. The physics-based set is COSMO-RS (COSMOtherm 2025) and COSMO-SAC (JCOSMO 3.1) with def-BP-TZVP and def2-BP-TZVPD parameterisations; the data-driven model is a temperature-dependent graph neural network (GNN) on molecular graphs. All are evaluated on the same database with MAE and RMSE in ln γ∞, R², and the Spearman rank correlation ρ, which governs screening. COSMO models predict activity coefficients qualitatively well but with large quantitative deviations in IL systems, regardless of model or level of theory; raising the level from TZVP to TZVPD-FINE unexpectedly increased COSMO-RS deviations. COSMO-SAC 25 and COSMO-RS (TZVP) presented the lowest deviations among COSMO models. The main difference is cost: 55k γ∞ took <2 min in JCOSMO versus about a day in COSMOtherm 2025 — three orders of magnitude. The GNN is far more accurate and ranks better still (ρ = 0.98 and 0.97) within its training domain, but degrades outside it, where no thermodynamic constraint applies. The result is a division of labour, not a winner: the GNN ranks in-domain at negligible cost, while COSMO gives physically constrained extrapolation to unseen chemistries. Within ILIMITED, it ranked a 19.5 M IL space for in-situ methanol/water sorption at 240 °C, and the best candidates are indeed in agreement with the experimental evaluation of literature data.