Thermoelectric devices, which transform heat flow into electrical power and vice versa, have the potential to revolutionize how society produces electricity and cooling. However, thermoelectric materials suffer from poor power conversion efficiency and the search continues for new materials with enhanced performance. In this project, advances in computation and machine learning are leveraged to accelerate this search for advanced thermoelectric materials. These efforts build upon the prior NSF DMREF research of some of team members on predicting a material's potential for thermoelectric performance. High throughput screening focused on identifying semiconductors with desirable electronic and vibrational properties. However, these efforts did not include a strong focus on the role of intrinsic defect or the potential for dopability. In the next stage of this research, these critical components will be pursued through a mixture of high throughput theory, experimental validation, and machine learning. Together, these efforts will yield accurate prediction of the thermoelectric potential for thousands of semiconductors and the realization of new materials for solid state power generation. Beyond thermoelectric materials, these efforts to establish a dopability recommendation engine will be critical in the development of next generation microelectronic and optoelectronic materials such as transparent conductors and photovoltaic absorbers.
's ultimate objective is to build a robust and accurate dopability recommendation engine to overcome the dopability bottleneck in thermoelectric materials discovery. The recommendation engine will use materials informatics to enable high-throughput predictions of dopability, relying only on quantities that are inexpensive to calculate, experimental measurements, and known structural/chemical features as inputs. It will thus allow dopability screening of thousands of compounds. First, an accurate training set will be built for the recommendation engine containing native defect formation enthalpies and structural/chemical descriptors from a diverse array of thermoelectric-relevant compounds. Whereas prior dopant studies focused on single compounds, a new, automated calculation infrastructure will be leveraged that allows the rapid creation of an extensive training set, initially containing approximately 30 compounds but growing to over 100 during the project. Experimental charge transport and local dopant structure measurements will validate the training set. Second, the prediction engine will be trained on the data to extract patterns and correlations, and ultimately identify robust descriptors of dopability. Initially, the engine will predict if `killer' defects limit the available dopant range. The engine will ultimately grow to suggest specific extrinsic dopants for compounds that pass this initial screening. Together, this combination of accurate predictions of intrinsic transport properties (prior DMREF) and dopability (proposed DMREF) is expected to accelerate the discovery process for thermoelectric materials.
