Figure 2
The effect of including experimental modifications in testing data on models trained without including any experimental modifications. Results are shown for unit-cell length predictions for (a) the cubic and (b) the triclinic crystal systems. The modifications studied correspond to zero error, intensity modulation, Gaussian broadening, Gaussian baseline noise and multiple impurity phases. Perfect refers to unmodified data and null refers to prediction using the mean lattice parameters of the data set. Zero error and intensity modulation have little effect on ML prediction. On the other hand, baseline noise and multiple phases are particularly damaging modifications. |