Table 5 The comparison of the performance of fine-tuned (ft) and pretrained SevenNet and MACE uMLIPs on the “w/o TM” and “with TM” parts of the nebDFT2k dataset

From: Benchmarking machine learning models for predicting lithium ion migration

MAE, eV

RMSE, eV

Rp

Slope

R2

Model

w/o TM

0.08

0.12

0.99

0.92

0.97

SevenNet

0.08

0.12

0.99

0.93

0.97

SevenNet-ft

0.07

0.12

0.99

0.96

0.97

MACE

0.10

0.13

0.99

0.91

0.96

MACE-ft

with TM

0.10

0.18

0.97

0.90

0.94

SevenNet-ft

0.13

0.25

0.94

0.86

0.88

MACE-ft

0.17

0.28

0.92

0.93

0.85

MACE

0.18

0.29

0.92

0.87

0.84

SevenNet

  1. The trajectories for which at least one interatomic potential fails optimization task were discarded. The metrics are calculated for 488 and 690 systems corresponding to “with TM” and “w/o TM” parts of the nebDFT2k dataset. Rows in the table are sorted ascending by RMSE for each subset of the data.
Back to article page