Table 2 Accuracy and speed for periodic systems

From: A Euclidean transformer for fast and stable machine learned force fields

 

ntrain

EMAE [\(\frac{\,\text{meV}\,}{{\rm{atom}}}\)]

FMAE [\(\frac{\,\text{meV}}{\mathring{\text{A}\,}}\)]

\(\frac{\mu \,\text{s}}{\text{step}\cdot \text{atom}\,}\)

Allegro53

10k

1.7

73.4

27.785*

SO3krates

10k

0.2

28.2

23.593*

SO3krates

1k

0.3

31.8

23.593*

  1. Speed in MD simulation and accuracy comparison to the strictly local Allegro model for Li3PO4 (192 atoms) on a single V100 GPU as reported in ref. 53. In ref. 53 no inference times and only MD step times with LAMMPS93 have been reported. This prohibits a purely model based comparison. We run MD simulations using the MDX code92, such that timings should be understood as illustration of the competitive nature in speed rather than an exact comparison.
Back to article page