Fig. 3: GAME-Net training with the FG-dataset.
From: Fast evaluation of the adsorption energy of organic molecules on metals via graph neural networks
a, FG-dataset illustration, showing the included chemical families, metals and surface facets with the corresponding crystal system. fcc, face-centered cubic; bcc, body-centered cubic; hcp, hexagonal closed packed. b, Bar plot of the adsorbate atom count in the FG-dataset. c, Distribution of the DFT energy target EA/M − EM. d,e, Box plot of the error distribution (d) and the MAE (e) grouped by chemical family in the test sets from the fivefold nested cross-validation. The colors in c–e are associated with the chemical families in a. The number of data used for d and e is n = 11,412, due to the fact that the k-fold nested cross-validation involves including all the graph data of the FG-dataset (n = 2,853) in the test set k − 1 times, each with a different combination of training and validation sets for training the model. Each box plot in d defines the median as the box center, the interquartile range (IQR) as the box size, with whiskers extending for 1.5 × IQR. Data in e are presented as mean ± s.e.m.
