Abstract/Summary

Machine learning is increasingly utilized to accelerate the discovery and design of new materials. Thermoelectrics are an important class of energy materials with the potential to help address pressing environmental challenges. This thesis presents novel machine learning-based methodologies for predicting material properties and generating crystal structures, with a focus on the discovery of new, and more effective, thermoelectric materials. First, a method is introduced for deriving distributed representations of materials solely from their chemical formulas, which demonstrates competitive performance in predicting various properties, such as formation energy and band gap. Next, an attention-based deep learning model is developed to predict thermoelectric transport properties, which incorporates the distributed representations, and proves capable of making useful predictions with a significantly reduced computational cost compared to traditional ab initio methods. Finally, a generative model is proposed that is capable of suggesting crystal structures for chemical compositions, which is vital for progressing from estimates of thermoelectric performance from composition, to deeper investigation based on structure. The results from these studies demonstrate the potential for modern machine learning techniques in the field of materials discovery, and particularly for accelerating the discovery of novel thermoelectrics.