fig3
Figure 3. Distribution of the shielding constants predicted by Lasso, and scatter plots of the shielding constants predicted by Lasso vs. those calculated by DFT for the test set Na40. To handle shielding constants that follow a generalized logistic distribution, in (B) and (D), the shielding constants are once transformed to a normal distribution to train the Lasso model. When making predictions, the prediction values of the Lasso model were back-transformed into shielding constants based on the original generalized logistic distribution. On the other hand, in (A) and (C), the shielding constants are predicted directly. In terms of the explanatory variables, (A) and (B) use the coordination numbers calculated from the structure, and (C) and (D) use the gross orbital population calculated by DFT calculations.