Fig. 1: Testing of MLIPs.
From: Discrepancies and error evaluation metrics for machine learning interatomic potentials
The comparison of the atomic environments from (a) the interstitial-RE testing set \({\mathcal{D}}\)RE-ITesting (blue), (b) the vacancy-RE testing set \({\mathcal{D}}\)RE-VTesting (red), and (c) the original training dataset from ref. 1 (cyan), using 6500, 6300, and 13553 atomic environments respectively. All atomic environments here are quantified by the 1st and the 2nd principal components from principal component analysis (PCA) of the SOAP descriptors. Comparison of the (g, h) atomic forces and (d, e) energies predicted from MLIP (GAP) versus the benchmark (DFT K4) on (e, h) the vacancy-RE testing set \({\mathcal{D}}\)RE-VTesting and (d, g) the interstitial-RE testing set \({\mathcal{D}}\)RE-ITesting. Phonon dispersion of (f) the bulk Si and (i) the Si supercell with a single vacancy calculated by DFT K4 and GAP.
