Skip to main content
. 2021 Jun 2;11:11604. doi: 10.1038/s41598-021-91068-8

Figure 2.

Figure 2

Illustration of the PCA-CGCNN architecture. (a) Dimension reduction of DOS vector by principal component analysis (PCA). The coefficients (α) became signal vector. (b) Construction of the crystal graph (CG) of NP structures and the structure of the convolutional neural network (CNN) on top of the CG. NP structures are converted to graphs with nodes and edges representing atoms and bonds, respectively. Then, the CNN processes are followed to reflect the local environments of each node in the CG. (c) ) Determination of signal vectors. After the CGCNN process, the new graph vector for each atom is fully connected with a signal vector for the DOS representation of each atom by neural networks. (d) DOS representation. With the signal vector obtained from the CGCNN, atomic DOS patterns are reconstructed on the basis of PCA. The sum of each atomic DOS pattern produces a total DOS pattern of the NP.