Active Learning to Accelerate the Search for New Materials with Emphasis on Domain Knowledge

Dezhen XUE*

 State key Laboratory for Mechanical Behaviors of Materials, Xi’an Jiaotong University, Xi’an, 710049

ABSTRACT: Informatics tools accelerate the search and discovery of new materials. Active learning, which iteratively augments informative samples, allows us to effectively navigate the huge search space. However, insufficient samples in materials science hinder the very predictive machine learning model. In addition to data, a distinguishing aspect of materials science is that there exists a substantial body of knowledge in the form of phenomenological models and physical theories.

We combined active learning and materials knowledge to design new materials descriptors, to establish better mapping between targeted properties and materials descriptors, and to effectively reduce the size of the unexplored material search space. We demonstrated the advantages of domain knowledge in developing high-performance lead-free ferroelectric ceramics. Firstly, a new descriptor describing the phase transition of ferroelectric materials is proposed, thereby improving the accuracy of the material performance regression model and classification model. Secondly, the "domain knowledge" is employed to preselect the crossover compositions between relaxor and normal ferroelectrics to narrow down the search space efficiently. Finally, based on a large amount of cheap data generated by the micromagnetism model, the mapping between microstructure and materials property is established to predict the property from experimental image. The above results show that the strategy of combining "domain knowledge" with active learning may be more efficient than pure data-driven research strategies.

Figure 1.  Active learning iteration loop with domain knowledge as a filer

Keyword: active learning; domain knowledge; descriptor; CNN