Fig. 5. Analysis of relevant interatomic features.
a Distributions of the average distance of pairwise features and average contribution of features to the force prediction for Ac-Ala3-NHMe, AT-AT, and the buckyball catcher using bivariate kernel density estimate plots (Machine Learning models: global - green, reduced before retraining - gray, reduced after retraining - green). The marginal charts on the top and right show the distribution of the two variables using density plot. The average values were obtained from all configurations in the datasets. The x axis is in log scale normalized to the 0–1 range. b Decomposition of the reduced descriptor by short- and long-range features for Ac-Ala3-NHMe, the AT-AT, and the buckyball catcher. Pairwise features with an average distance below 5 Å across all configurations in the dataset are counted as short-range (green line), long-range otherwise (blue line). Dashed green line represent number of short-range features in the global descriptor. Descriptor sizes in x axis go from 1 to 0, where 1 corresponds to a default global descriptor and 0 to an empty descriptor.