Fig. 8: ML-predicted and DFT-calculated dielectric constants for the candidate materials. | npj Computational Materials

Fig. 8: ML-predicted and DFT-calculated dielectric constants for the candidate materials.

From: Dielectric tensor prediction for inorganic materials using latent information from preferred potential

Fig. 8

The joint distribution of band gaps Eg and electronic dielectric constants ε∞ (a) or ionic dielectric constants ε0 (b) for 6648 training structures and 14,375 stable candidate structures. ε∞ and ε0 of training structures and Eg of all structures are obtained from computational results in MP, while ε∞ and ε0 of candidate structures are predicted by DTNet-\({\varepsilon }^{\infty }/{\varepsilon }_{}^{0}\). The 5-level densities of data points are drawn in contours using the kernel density estimation for structures both in the candidate set and original set. c DFT-calculated Eg − ε data points colored by their figure of merit (FOM) defined by Eg ⋅ ε for 4146 stable structures in training data and 35 new stable structures screened by DTNet across 3 iterations. d Structure visualization of the top 3 materials by FOM as validated by DFPT calculations.

Back to article page