Table 1 Benchmarking pre-trained extractor aggregation methods.

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)

Ensemble

0.221 ± 0.034

48.8 ± 10.0

0.0951 ± 0.0054

Concatenate

0.238 ± 0.058

54.2 ± 12.9

0.108 ± 0.007

Add

0.220 ± 0.029

48.1 ± 9.0

0.0982 ± 0.0080

  1. Average test mean absolute errors (MAE) and standard deviations over five random seeds for different methods of aggregating pre-trained extractors. Smaller MAEs are better. Each method was benchmarked on predicting piezoelectric modulus (d33)34, 2D exfoliation energies (Eexfol)35, and experimental formation energies (Expt. Ef)36,37.
Back to article page