Fig. 6.

Activation analysis to understand the impact of transfer learning. Here, we analyze the activations from the first hidden layer of the ElemNet architecture for understanding the impact of transfer learning on the model’s capability to automatically learn to distinguish between the magnetic vs non-magnetic class (1 and 0) from JARVIS data set. The four columns represent the models trained using four different data sets: a, e, and i using JARVIS (JAR), b, f, and j using Materials Project (MP), c, g, and k using OQMD and d, h, and l using the experimental observations (EXP); the first a–d and second e–h (except g) rows represent scatter plots demonstrating the first two principal components of the activations using principal component analysis (PCA) technique from the models trained from scratch (SC) and using transfer learning (TL), whereas third row i–l represents the ROC curves from the Logistic Regression model trained using complete set of activations from the same hidden layer (the corresponding AUC values are shown in brackets) on the corresponding data sets.