Fig. 4: Evaluation of the validity of generated materials. | npj Computational Materials

Fig. 4: Evaluation of the validity of generated materials.

From: Generative adversarial networks (GAN) based efficient sampling of chemical composition space for inverse design of inorganic materials

Fig. 4

a The percentages of charge-neutral (CN) and electronegativity-balanced (EN) samples of the generated samples are very close to those of the training sets for all four datasets. Train/gen CN: percentage of training/generated samples that satisfy charge neutrality; Train/gen EN: percentage of samples that satisfy balanced electronegativity. b Formation energy distribution of the Li-containing compounds generated by three GANs. Both GAN-ICSD and GAN-MP can generate a large percentage of hypothetical materials with low (<0) formation energy.

Back to article page