Table 2 Comparing MoE with the best-performing pairwise TL model.

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

 

d33

Eexfol (meV/at)

Expt. Ef (eV/at)

STL

0.228 ± 0.033

62.0 ± 13.6

0.194 ± 0.027

Best TL-(3)

0.222 ± 0.022a

52.7 ± 9. 0b

0.117 ± 0.007c

Best TL-(18)

0.215 ± 0.024

51.7 ± 7.6

0.117 ± 0.007

MoE-(3)

0.220 ± 0.029

48.1 ± 9.0

0.0982 ± 0.0080

MoE-(18)

0.206 ± 0.026

52.8 ± 10.5

0.0946 ± 0.0054

  1. aTL from MP bandgaps outperformed TL from MP formation energies and MP bulk moduli with MAEs of 0.276 ± 0.091 and 0.229 ± 0.020, respectively.
  2. bTL from MP formation energies outperformed TL from JARVIS formation energies and MP shear moduli with MAEs of 53.9 ± 11.6 and 59.4 ± 12.4, respectively.
  3. cTL from MP formation energies outperformed TL from JARVIS and perovskite formation energies with MAEs of 0.119 ± 0.008 and 0.289 ± 0.024, respectively.
  4. Average test MAEs and standard deviations over five random seeds for single-task learning, pairwise transfer learning, and mixture of experts using three human-chosen and all 18 available pre-training tasks. Best MAEs are bolded. See Table 3 for downstream data source references and Supplementary Table 2 for details regarding pre-training tasks.
Back to article page