Fig. 2: Schematics and architecture of AdsGT encoder designed for surface graphs.
From: A multi-modal transformer for predicting global minimum adsorption energy
a The positional encoding in AdsGT encoder. AdsGT encoder uses positional encoding to represent the relative height of each atom on the surface, where importance increases from bottom to top. Each atom i is assigned a positional feature \({\delta }_{{{{\rm{i}}}}}={H}_{{{{\rm{i}}}}}/{H}_{\max }\), where Hi denotes the atom’s height and \({H}_{\max }\) is the maximum height in the surface. b The architecture of AdsGT encoder. AdsGT encoder includes radial basis function (RBF) expansions, embeddings, and graph attention layers. c Illustration of each AdsGT layer with an edge-wise attention mechanism delineated by three steps: calculation of edge-wise attention coefficients, edge-wise message calculation, and node update. \({d}_{ij}^{t}\) and \({e}_{ij}^{t}\) are the distance and embedding of t-th edge between atom i and j. zi is the atomic number of atom i. \({h}_{i}^{l}\) is the atomic embedding of atom i at l-th AdsGT layer. \({{{\rm{LN}}}}\) and BN represent layer normalization and batch normalization, respectively.
