Fig. 5: Adsorption site identification by cross-attention scores.
From: A multi-modal transformer for predicting global minimum adsorption energy
a Schematic of identifying the most energetically favorable adsorption sites from the average cross-attention score of each surface atom relative to the adsorbate, which is calculated over all attention heads in the last cross-attention layer. b The accuracy (n = 5) of AdsMT models adopting different graph encoders48,49,50 in identifying optimal adsorption sites with or without transfer learning (TL). The black dotted lines represent the accuracy of random atom selection. The error bars represent standard deviations from five experiments. c Four examples of the comparison between (left) global minimum adsorption structures optimized by density functional theory (DFT) and (right) attention score-colored surfaces computed by the trained AdsMT model with AdsGT encoder. The black arrows point to the most stable adsorption sites.
