fig6

Figure 6. (A) CNN architecture for band structure prediction; (B) Tensor representation and physical insights incorporated into the CNN model; (C) Overall architecture of the TGNN and the crystal property prediction process based on TGNN; (D) a: Traditional band gap calculation using the GW approximation. b: Band gap estimation using the TGNN model; (E) Comparison between PBE-GGA (blue squares) and predicted G₀W₀ band gaps (red circles)^[198,199]. CNN: Convolutional neural network; TGNN: tuple graph neural network; GW: GW approximation; PBE-GGA: Perdew-Burke-Ernzerhof generalized gradient approximation.