Fig. 5.

Prediction error analysis on the experimental data set. The experimental data set contains 1,963 observations. For the models trained using experimental data set, the predictions are aggregated on test (validation) sets from each split in the 10-fold cross-validation. For the models trained using JARVIS and Materials Project, since we have 10 models from the 10-fold cross-validation during training, we take the mean of their predictions for each data point in the experimental data set. For OQMD-SC, we make 10 predictions for each point in the experimental data set and take their mean. The four rows represent the four data sets: a–c JARVIS (JAR), d–f Materials Project (MP), g–i OQMD, and j–l the experimental observations (EXP); first a, d, g, and j and second b, e and k (except h) columns of each row show the predictions using the model trained on the particular data set from scratch (SC) and using transfer learning (TL) respectively; the third column c, f, i, and l shows the corresponding CDF of the prediction errors using models trained from scratch (SC) and using transfer learning (TL).