(a) Test error evaluated on an unknown
test set containing 3840
APTs stemming from 10 randomly sampled MD snapshots as a function
of training data set size (learning curve, top panel). Mind the log–log
scale of the plot and that the expected linear relation (indicated
by the red dashed line) is recovered. In the bottom panel, the MSE
on the respective validation set is shown as a function of training
epoch and training data set size, i.e. four (black), 9 (brown), 18
(red), and 27 (green) MD snapshots, corresponding to 1536, 3456, 6912,
and 10368 APTs, respectively. (b) Performance of the APTNN trained
on 3840 APTs (9 MD snapshots) on an unknown test set containing 3840
APTs stemming from 10 randomly sampled MD snapshots. The figure compares
all components of the APT matrix individually for O (blue) and H (red)
atoms. Note that the test set used to benchmark the APTNN in the top
panel of (a) and in (b) is the same.