Multi-task Learning and Data Augmentation for Negative Thermal Expansion Materials Property Prediction

ABSTRACT Increasingly, data-driven materials informatics is regarded as a powerful means for new materials research and development. However, small imbalanced data and weak generalizability learning algorithm are two main challenges from the data and model view of points. In this paper, we propose a multi-task learning algorithm with augmentation data preprocessing to deal with the small imbalanced data and multi-target predictions. For experiments, we use negative thermal expansion materials as an example. Three targets, starting temperature T1, ending temperature T2, and the negative expansion coefficient α are prediction targets. The experimental results indicate that, firstly, after data augmentation, the prediction accuracy of all three targets are improved. Moreover, a few rounds of data augmentation can improve the prediction accuracy, but with more augmentation runs, it results in accuracy decreasing. Secondly, comparing with support vector regression and random forest, multi-task learning has achieved the best prediction accuracy in T1, T2, and α. The detailed R2 of T1, T2, and α are 0.8224, 0.8795 and 0.8984 respectively, which indicates multi-task learning a great potential in broad applications in future.