Fig. 5: Parity plots and error distribution for two control studies. | npj Computational Materials

Fig. 5: Parity plots and error distribution for two control studies.

From: A database of experimentally measured lithium solid electrolyte conductivities evaluated with machine learning

Fig. 5

a The shuffled control parity plot demonstrates each materials actual conductivity plotted against an average of five randomly selected values across the dataset. c The distribution of errors across all experiments (without averaging) demonstrates the maximal error bounds we would expect from a poor statistical model, with 68% of predictions falling between −2.36 to 2.31 away from the true values. b The mean control experiment demonstrates the expected predictions for a model which has simply learnt the mean value of the dataset. Correspondingly, the distribution of errors (d) is simply a reflection of the distribution of conductivities around the mean value, and models which form predictions close to the mean will resemble this distribution. A Student’s t-distribution (orange) is fit to the underlying data, with the mean of this distribution (dark blue), and the first, second, and third standard deviations away from this mean (light blue) overlaid in (c) and (d). A good model should have a mean of zero, with tight error bounds.

Back to article page