Prediction of material property using optimized augmented graph-attention layer in GNN

Research in materials science continues to focus on the development of ML models capable of accurately predicting material properties (MPP). Graph neural networks (GNNs) have also proven to be quite actual in extracting physicochemical properties from MPP, despite the fact that many other techniques have been offered in the past. Despite this, existing GNN models fail to distinguish between the contributions of various atoms adequately. To forecast the properties of inorganic materials, we have developed an upgraded graph neural network model dubbed IGAT-GNN. OAGAT and a global attention layer, where a meta-heuristic method is used to achieve optimal weight in OAGAT, are the distinguishing features of IGAT-GNN. Each of the two methods, OAGAT layers and global attention layers, is used to better understand the local relationships among nearby atoms and their overall influence to the material's stuff. This results in much improved prediction performance for various qualities that have been examined. As we have demonstrated through numerous experiments, the IGAT-GNN outperforms other currently available GNN models while also revealing new information about the relationship between atoms and their respective material properties. The proposed model alewives the accuracy of 99.95%, which is the high quality accuracy than other algorithm. © 2022