Fig. 2: Performance evaluation of UNEP-v1 using the training and test datasets.
From: General-purpose machine-learned potential for 16 elemental metals and their alloys
a–c Parity plots for energy, stress, and force comparing density functional theory (DFT) reference data and the first version of unified neuroevolution potential (UNEP-v1) predictions for the whole training dataset. In c there are three test datasets containing n-component (n ≥ 3) structures, including one with up to 13 components (Ag, Au, Cr, Cu, Mo, Ni, Pd, Pt, Ta, Ti, V, W, Zr) taken from Lopanitsyna et al.38 (labeled Test-1), one with up to four components (Mo, Ta, V, W) from Byggmästar et al.37 (labeled Test-2), and one with up to three components (Pd, Cu, Ni) from Zhao et al.36 (labeled Test-3). d, e Parity plots for formation energies comparing DFT reference data and predictions from UNEP-v1 (green circles), MACE-MP-0 (medium model, blue stars)40, and embedded-atom method (EAM) (orange triangles)35, for structures from the Materials Project (MP-ternary)33 and the GNoME paper39. Mean absolute error (MAE) and R2 (coefficient of determination) values are provided for comparison. f Distribution of the training dataset (this work, UNEP-v1, comprising 1-component to 2-component systems, blue) and various test datasets, including Test-1 (up to 13-component systems, orange)38, Test-2 (up to 4-component systems, yellow)37, Test-3 (up to 3-component systems, purple)36, MP-ternary alloys (3-component systems, red)33, and GNoME dataset (2-component to 5-component systems, green)39, in the 2D principal component (PC) space of the descriptor. Source data are provided as a Source Data file88.
