Table 1 Energy and force RMSEs for six different NN architectures.

From: Unfolding the structural stability of nanoalloys via symmetry-constrained genetic algorithm and neural network potential

NN architecture

\({E}_{{{{\rm{train}}}}}^{{{{\rm{RMSE}}}}}\)

\({E}_{{{{\rm{test}}}}}^{{{{\rm{RMSE}}}}}\)

\({F}_{{{{\rm{train}}}}}^{{{{\rm{RMSE}}}}}\)

\({F}_{{{{\rm{test}}}}}^{{{{\rm{RMSE}}}}}\)

10-10 (p-p)

4.46

4.41

75.63

75.80

10-10-10 (p-p-p)

2.00

2.20

75.59

75.75

15-15 (p-p)

1.72

1.96

71.25

71.31

15-15 (s-s)

1.70

1.85

72.06

72.19

15-15 (t-t)

1.67

1.84

71.79

71.94

15-15-15 (p-p-p)

1.71

1.95

71.05

71.37

20-20 (p-p)

1.54

1.77

69.07

69.40

20-20-20 (p-p-p)

1.58

1.81

68.52

69.06

  1. The architectures are denoted by the number of neurons and the activation functions (p = softplus, s = sigmoid, t = hyperbolic tangent) used in the hidden layers. The RMSEs are obtained from the training and test sets after 30 training epochs. The units for the energy and force RMSEs are meV ⋅ atom−1 and meV ⋅ Å−1 ⋅ atom−1, respectively.
Back to article page