Performance of Neuroevolution Potentials for the Ionic Liquid [Emim][Ac]: A Systematic Benchmark of Four Density Functionals

Abstract Machine-learned force fields (MLFFs) based on Neuroevolution Potential (NEP) models were developed for the ionic liquid 1-ethyl-3-methylimidazolium acetate ([Emim][Ac]) using first-principles reference data generated with four density functionals: PBE-D3, RPBE-D3, optB88-vdW, and SCAN+rVV10. Using a unified training protocol, the resulting NEP models were systematically benchmarked against structural, thermophysical, transport, and hydrogen-bond-network properties, including radial distribution functions (RDFs), coordination numbers, neutron total scattering profiles, density, viscosity, ionic conductivity, and thermal conductivity. Spearman correlation analysis and weighted composite scoring were further used to establish a property-oriented evaluation framework for MLFF assessment and selection. The results reveal a pronounced dependence of NEP performance on the density functional used to generate the reference data. Although all four models reproduce the dominant local structural features of [Emim][Ac], appreciable differences emerge in medium-range structural ordering and transport properties. The PBE-D3-based NEP provides the most accurate overall description of structural properties, whereas the optB88-vdW-based NEP yields the closest agreement with experimental transport properties, particularly viscosity and ionic conductivity. Analysis of the hydrogen-bond network further reveals systematic relationships among hydrogen-bond populations, residence times, free-ion fractions, and ion transport. These results demonstrate that, for [Emim][Ac], the exchange–correlation functional used for reference-data generation influences the physical fidelity and property-specific performance of the resulting NEP models.