Fig. 2: An overview of the nebBVSE122k and nebDFT2k datasets and results of benchmarking the selected models for structure-to-property prediction of Li-ion migration barriers.

a, b An illustrative example of linearly interpolated and BVSE-NEB optimized Li-ion migration trajectory, respectively, and c corresponding energy profile, d Distributions of the BVSE- and DFT-NEB calculated Li-ion migration barriers, e Parity plot for the test data predictions obtained with the M3GNet model trained on the nebBVSE122k, f The correlation between the weight of the edge connected to the centroid and its length extracted from the last layer of the Allegro model trained on the nebBVSE122k dataset, g An illustrative example of the structure with the centroid dummy element, h Parity plots for the test data predictions obtained with the SVR model trained on the nebDFT2k datasets, i The top-5 most important features according to the impurity-based method (f1 – minimum volume of the Voronoi polyhedra constructed around the initial, final, and centroid Li sites, f2 – edge length, f3 – maximum weighted average of oxidation states of the nearest neighbors collected for initial, final, and centroid Li sites, f4 – maximum volume of the Voronoi polyhedra constructed around the initial, final, and centroid Li sites, f5 – minimal average covalent radius of nearest neighbors collected for initial, final, and centroid sites.