Materials By Design: A Combined Computational and Experimental Approach to Thermoelectrics

Miller, Samuel A

doi:https://doi.org/10.21985/n2-tptg-mk88

Work

Materials By Design: A Combined Computational and Experimental Approach to Thermoelectrics

Public

Download PDF

Download All Files (.zip)

In the United States and around the world, the growing energy demands and climate concerns necessitate renewable and efficient energy production. Thermoelectric materials could be one small part of this larger picture movement, but their high cost and low efficiency must be improved to realize commercial use. To decrease the time to application of new compounds, computations which provide predictions of material properties may guide the discovery and investigation of novel thermoelectric materials. For these models to be useful for high-throughput screening of numerous compounds across a varied phase space, they must be accurate enough to provide trustworthy predictions and inexpensive enough to be suitable for large and diverse groups of compounds. This work focused on the development of empirical models for predicting material properties relevant to thermoelectric applications as well as the use of these models to guide experimental investigation of interesting compounds. A semi-empirical model was constructed by fitting physics-based equations to experimental properties using inexpensive inputs. By utilizing inputs which are readily accessible from known structural data and basic first-principles calculations, this model is capable of being implemented for high-throughput predictions without limitations on the complexity or number of compounds considered. Since we begin with analytical expressions for these properties and utilize a small number of fitting parameters, rationalization of the effects and outputs is straightforward. This model was refined by adding additional data to the learning set, introducing a structural parameter, and statistical fitting, thereby increasing the accuracy. SnO, GeSe and ZrTe5 are three layered compounds which were investigated for their thermoelectric performance. While all exhibit high potential based on predictions using the semi-empirical model, realization of this potential through optimization proved difficult. Individually, each compound poses a unique challenge; collectively, they demonstrate that chasing high potential alone is not recommended. Rather the possibility of realizing the optimum performance must be acknowledged. If we are to use predictions from models to guide experiments, we must predict dopability in addition to thermoelectric potential. Dopability estimates are accomplished through statistical and machine learning on experimental carrier concentration data gathered from the literature. Predictions using this model are accurate to approximately one order of magnitude and the drivers of dopability are rationalized and compared to previous efforts due to the simplicity of the model. The combination of these two models, one for optimum performance and the other for dopability, demonstrates the effectiveness of building prediction engines from empirical data to allow human learning and guide further experimental and computational work.

Creator