Fig. 3: Performance of DeepH on studying graphene. | Nature Computational Science

Fig. 3: Performance of DeepH on studying graphene.

From: Deep-learning density functional theory Hamiltonian for efficient ab initio electronic-structure calculation

Fig. 3: Performance of DeepH on studying graphene.

a, MAE of \({H}_{i\alpha ,\,j\beta }^{\prime}\) for different orbitals. b, Distribution of \({H}_{i1,\,j1}^{\prime}\) for the nearest neighbors (atomic distance between 1.28 and 1.6 Å; see the corresponding distribution in the inset). The s.d. of the computed \({H}_{i1,\,j1}^{\prime}\) is 315 meV for the test set. c,d, Distribution of the generalization MAE of the DOS for 2,000 unseen material structures (c). Three typical structures with the best, median and worst MAE values for the DOS (atomic structures included in Supplementary Data 1) are indicated. Their DOS (c, inset) and shift current conductivity σyyy (d), computed by DFT and DeepH, are compared.

Source data

Back to article page