Data-driven thermoelectric modeling: Current challenges and prospects

Abstract

Recent advancements in computing technologies coupled with the need to make sense of large amounts of raw data have renewed much interest in data-driven materials design and discovery. Traditional materials science research relies heavily on experimental data to gauge the properties of materials. However, this paradigm is purely based on trial and error and ongoing research can take decades to discover new materials. Data-driven modeling tools such as machine learning and its proven libraries can help speed up the materials’ discovery process through the implementation of powerful algorithms on readily available material datasets mined from the ever-increasing private- and government-funded material databases. In this Perspective, we applied various machine learning models on tens of hundreds of thermoelectric compounds obtained from density functional theory calculation results. In our preliminary analysis, we made use of pymatgen and the powerful materials science library matminer to add and explore key material features that have the propensity to accurately predict our achievable target output. We evaluated the accuracy and performance of our models with the coefficient of determination (R2), the root mean square error, and K-fold cross-validation metrics and identified the most important descriptors for our materials. Finally, we reviewed the current state-of-the-art in data-driven thermoelectric materials’ design and discovery, its current challenges, and prospects.