455131 Machine Learning for Advancing Discovery of Novel Thermoelectric Materials. the Thermoel

Monday, November 14, 2016: 2:12 PM
Yosemite A (Hilton San Francisco Union Square)
Al'ona Furmanchuk1, Ankit Agrawal1, James Saal2, Jeff Doak2, Gregory Olson2 and Alok Choudhary1, (1)Northwestern university, Evanston, IL, (2)QuesTek Innovations LLC, Evanston, IL

Based on current estimates 59% of all primary energy produced in the United States is lost as waste heat, recovery of even 1% of this heat by a novel thermoelectric materials would amount to nine times the energy currently produced by photovoltaics. Thermoelectric materials are materials that exhibit a strong coupling between electric and thermal responses, producing an electric potential in the presence of a temperature gradient and vice versa.
The physics governing the material systems is described as a network of models and functional relationships between different scales and modalities (i.e. processing/structure/properties). However, the disjointed nature of materials property databases and relative immaturity of the thermoelectrics multi-scale functional relationships make integrated optimization across numerous length and time scales particularly difficult in this case. A data-driven design of advanced thermoelectrics is possible with tools capable of integrating materials databases into a consistent framework and finding the multi-scale underlying processing/structure/property functional relationships responsible for thermoelectrics performance.
The ThermoEl could be an effective solution to discovery of novel thermoelectrics. It is a predictive toolkit developed based on a supervised machine learning algorithms and known experimental measurements of various materials. Aside from other tools on the market, ThermoEl is based on regression models that predict numerical value of such parameters as Seebeck coefficient for non-stoichiometric materials and Bulk modulus of crystalline compounds. Based on our experience, we would like to discuss tricks and challenges in using data mining approaches with materials databases.

Extended Abstract: File Not Uploaded