fig5
Figure 5. Distribution of shielding constants predicted by the neural network and the scatter plot of the shielding constants predicted by the neural network vs. the shielding constant calculated using the DFT based on the test set Na40. Two approaches were employed for handling the generalized logistic distribution of shielding constants. One approach is to convert the shielding constant to a normal distribution before training the neural network in (C) and (D). However, before plotting, the neural network predictions are converted to the original generalized logistic distribution. In contrast, in (A) and (B), the shielding constants are learned directly. The other approach used Equation (1) as the loss function in (B) and (D). In contrast, (A) and (C) employ only MSE as the loss function.
