Table 3 Performance of the target-oriented MLP models compared with other potentials

From: Transferable machine learning model for multi-target nanoscale simulations in hydrogen-carbon system from crystal to amorphous

Target feature space

Empirical potentials

GAP

MLP (Best performance)

REBO

Tersoff

REBO-s

Tersoff-s

Reaxff

∆E

∆F

∆E

∆F

∆E

∆F

∆E

∆F

∆E

∆F

∆E

∆F

∆E

∆F

Deposition

1.114 (0.35)

12.71 (22.3)

—

—

1.320 (0.316)

27.33 (26.53)

—

—

0.127 (0.078)

7.64 (6.61)

—

—

0.019 (0.007)

0.78 (0.44)

Quench

0.990 (0.140)

5.84 (4.89)

1.617 (0.354)

6.97 (5.37)

1.719 (0.656)

18.35 (16.09)

2.186 (0.479)

7.88 (5.37)

0.403 (0.289)

6.07 (4.49)

0.198 (0.033)

0.68 (0.58)

0.021 (0.010)

0.93 (0.54)

Friction

0.816 (0.273)

2.81 (2.11)

1.159 (0.051)

5.33 (4.14)

0.921 (0.293)

3.76 (13.36)

1.502 (0.045)

7.48 (5.41)

0.138 (0.095)

5.10 (3.33)

0.135 (0.021)

4.43 (0.52)

0.017 (0.010)

0.42(0.29)

Crack-tip

0.649 (0.048)

2.03 (1.80)

0.757 (0.213)

2.93 (2.584)

0.606 (0.046)

2.98 (2.65)

0.775 (0.241)

3.33 (2.96)

0.315 (0.113)

4.56 (3.84)

0.093 (0.028)

1.38 (2.58)

0.034 (0.020)

0.54 (0.52)

  1. The energy (ΔE) and force (ΔF) root mean square error (RMSE) values compared with DFT calculations in eV/atom and eV/Å were collected on different test datasets at target feature space. The numbers in parentheses are the standard deviation of the ΔE and ΔF in three times random sampling of the test datasets. The bold numbers represented the highest accuracy of target-oriented MLP models corresponding to different target datasets.
Back to article page