Materials representation and transfer learning for multi-property prediction

The adoption of machine learning in materials science has rapidly transformed property prediction. Hurdles limiting full capitalization recent advancements include the limited development methods to learn underlying interactions multiple elements as well relationships among properties facilitate prediction new composition spaces. To address these issues, we introduce Hierarchical Correlation Learning for Multi-property Prediction (H-CLMP) framework that seamlessly integrates: (i) using only a material's composition, (ii) and exploitation correlations target multi-target regression, (iii) leveraging training data from tangential domains via generative transfer learning. model is demonstrated spectral optical absorption complex metal oxides spanning 69 three-cation oxide H-CLMP accurately predicts non-linear composition-property spaces which no are available, broadens purview discovery with exceptional properties. This achievement results principled integration latent embedding learning, correlation attention models. best performance obtained [H-CLMP(T)] wherein adversarial network trained on computational density states deployed domain augment composition. H-CLMP(T) aggregates knowledge sources suited regression across physical sciences.