DR-MHEA-GCN: A Semisupervised Graph Framework with Dual-Residual Edge Attention for Chemical Bond Classification in...

Identifying chemical bond types is fundamental to understanding structure-property relationships in two-dimensional (2D) nanomaterials, yet a unified bond-classification strategy remains lacking. Here, we develop a dual-residual multihead edge-attention graph convolutional network (DR-MHEA-GCN) for atomic-scale chemical bond identification in 2D systems. The network integrates dual-residual connections to stabilize deep feature propagation and a multihead edge attention mechanism to selectively emphasize bond-relevant interactions. Approximately 3,000 2D nanomaterials are encoded as atomistic graphs with physically motivated atomic and bond descriptors. By incorporating pseudolabel refinement and structural perturbation, DR-MHEA-GCN achieves 95.3% test accuracy and 98.1% agreement on extrapolative data, capturing intrinsic interatomic bonding characteristics while accelerating analysis by approximately 3 orders of magnitude over conventional DFT-based methods. Interpretable analysis further reveals that the model operates through specialized physical perspectives (i.e., distinct latent subspaces) and confirms that atomic electronegativity plays a dominant role in bond prediction, surpassing electronegativity difference in importance.