Fig. 5: Errors of atomic forces.
From: Discrepancies and error evaluation metrics for machine learning interatomic potentials
Illustrations of (a) force magnitude error \({\delta }_{{\rm{F}}}\) and (f) force direction \({\delta }_{{\rm{\theta }}}\) of atomic force Fpredicted predicted by an MLIP in comparison to FDFT by DFT. For the RE atoms in the \({\mathcal{D}}\)RE-ITesting, the errors (b) \({\delta }_{{\rm{F}}}\) and (g) \({\delta }_{{\rm{\theta }}}\) of DeePMD (orange) and GAPPRX (cyan), the corresponding distribution of (c) \({\delta }_{{\rm{F}}}\) and (h) \({\delta }_{{\rm{\theta }}}\), and the cumulative distribution function (CDF) of (d) \({\delta }_{{\rm{F}}}\) and (i) \({\delta }_{{\rm{\theta }}}\). The CDF of (e) \({\delta }_{{\rm{F}}}\) and (j) \({\delta }_{{\rm{\theta }}}\) on vacancy RE atoms in the \({\mathcal{D}}\)RE-VTesting set. Interstitial- and vacancy-RE atoms are atoms in the middle of migration identified in the \({\mathcal{D}}\)RE-ITesting and \({\mathcal{D}}\)RE-VTesting datasets, respectively.
