Data-Driven Design of Nanoscale Features to Obtain High-zT Thermoelectrics

We describe methods for improving efficiency in thermoelectric materials, which we have explored through a statistical learning-based investigation into design of optimal structural features. Structure, composition, dimensionality, and corresponding thermoelectric properties have been extracted from 57 different thermoelectric materials systems publications and compiled into a database. Feature selection methods such as ANOVA, lasso regression and PCA were used to refine the dataset and remove unimportant features that have negligible impact on zT. Afterwards, extra trees, nonlinear SVM, and Gaussian process regressors were built on the data set, with the efficacy of each evaluated by the estimated error via leave-one-out cross-validation. Finally, we will discuss optimal experimental design techniques, which have been implemented to verify the model and to exploit the design space such that criteria for high-zT thermoelectrics can be extracted.