High-throughput application and evaluation of the COSMO-SAC model for predictions of liquid–liquid equilibria

Abstract

The predictive power of the COSMO-SAC activity coefficient model is rigorously put to the test using an extensive dataset of binary liquid–liquid equilibria (LLE). Two model variants, COSMO-SAC-2010 and COSMO-SAC-dsp, are evaluated across 2478 binary systems and nearly 75 000 experimental data points. They achieve a success rate exceeding 90% in detecting the occurrence of LLE, demonstrating strong qualitative performance across chemically diverse systems. In benchmark comparisons, COSMO-SAC-2010 sets the standard for nonaqueous systems, while COSMO-RS performs best for aqueous mixtures, placing the two at a broadly comparable overall level with complementary strengths. COSMO-SAC-dsp yields larger deviations but provides broader coverage, particularly for polar and hydrogen-bonding systems. Both reliably capture systematic trends across homologous series, making them effective tools for solvent screening and thermodynamic consistency analysis. A high-throughput and fully automated computational framework—integrated into the open-source package ThermoSAC—enables adaptive Gibbs energy screening, LLE tracing, and anomaly detection. This work establishes COSMO-SAC as a leading framework for predictive thermodynamics and offers reproducible benchmarks and tools for future model development, such as those based on machine learning.