Fig. 1: Schematic of adversarial transfer learning from bulk effective mass to 2D carrier mobility.
a Adversarial transfer learning. Both bulk and 2D materials are first transformed into feature vectors based on their structures and compositions using materials materials-agnostic platform for informatics and exploration (MAGPIE), and then a multi-layer perceptron is used to extract features. The extracted features are used for two tasks: bulk effective mass regression and material classification. When backpropagating the regressor (R) loss and the reversed classifier (C) loss, the feature extractors are trained to extract only shared features between the target and source domains. E and C are trained iteratively until C can no longer identify bulk and 2D materials. b 2D carrier mobility prediction with expert knowledge. Using H-MoS2 as an example, features related to effective mass are extracted by E from its POSCAR, moreover, based on expert knowledge, features like mirror (M) and rotational (R) symmetry, layer thickness (T), and valence electron number (Ne) are also extracted. With this hybrid feature vector, final predictions on 2D carrier mobility are given by well-trained regression tree models.
