Fig. 4: Correlations between various similarity coefficients on the test set after training on 300 materials.
From: Machine learning climbs the Jacob’s Ladder of optoelectronic properties
The columns refer to SC[RPADL;RPADFT] a and SC[RPATL;RPADFT] b, while the rows refer to SC[IPADFT;RPADFT] c, SC[IPAML;IPADFT] d. The histogram in the first row shows that the ab initio IPA spectra are only a rough approximation of the ab initio RPA spectra. The histogram in the second row reproduces the observation of Ref. 18 that very good ML predictions of the DFT-IPA spectra are possible. The histograms in the top panels show that TL predicts the ab initio RPA spectra much better than DL. The two-dimensional histograms in the center show the correlations between the SCs defining columns and rows. All SCs are evaluated on the test set. The dashed orange line marks the angle bisector.
