Fig. 3: Performance of the trained SNAP model for ZIF-8. | npj Computational Materials

Fig. 3: Performance of the trained SNAP model for ZIF-8.

From: Quantum-accurate machine learning potentials for metal-organic frameworks using temperature driven active learning

Fig. 3

a Learning curves for the RMSE and MAE for energy (left-hand side panel), forces (middle panel) and virial-stress (right-hand side panel). Data are presented for test set A (composed of ~5000 configurations from AIMD simulations) and test set B (composed of ~2000 configurations from classical MD simulations) as a function of the number of configurations in the training set. Note that no significant change in the error is observed after the training set size reaches ~ 600 configurations. b Parity plots for energy, forces and virial-stress values comparing DFT and SNAP (trained over 672 configurations) values. The RMSE is 0.7 meV/atom, 86 meV/Å, and 29.5 MPa, respectively for energy, forces and virial-stress.

Back to article page