Table 4 MoE-(18) with sparse gating.

From: Towards overcoming data scarcity in materials science: unifying models and datasets with a mixture of experts framework

k

\({{d}_{33}}^{{\mathrm{a}}}\)

\({{E}_{{{{\rm{exfol}}}}}}^{{\mathrm{b}}}\)

Expt. \({{E}_{{{{\rm{f}}}}}}^{{\mathrm{c}}}\)

2

0.220 ± 0.029

54.0 ± 10.1

0.125 ± 0.047

4

0.217 ± 0.028

54.4 ± 11.8

0.101 ± 0.008

6

0.214 ± 0.022

55.3 ± 11.1

0.0990 ± 0.0047

10

0.208 ± 0.024

55.8 ± 10.2

0.0958 ± 0.0059

18

0.206 ± 0.026

52.8 ± 10.5

0.0946 ± 0.0054

  1. aPiezoelectric modulus from Matminer’s piezoelectric_tensor dataset34.
  2. bMonolayer exfoliation energies (meV/atom) from Matminer’s jarvis_dft_2d dataset3,35.
  3. cExperimental formation enthalpies (eV/atom) from Matminer’s expt_formation_enthalpy37 and expt_formation_enthalpy_kingsbury36 datasets. The former was preferred when duplicates arose.
  4. Average test MAEs and standard deviations over five random seeds for MoE with different settings of the sparsity hyperparameter k.
Back to article page