| About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Alloys and Compounds for Thermoelectric and Solar Cell Applications XIII
|
| Presentation Title |
New Efficient Half-Heusler Compositions from Machine Learning, DFT Calculations and Experiments |
| Author(s) |
Philippe Jund, Shoeb Athar |
| On-Site Speaker (Planned) |
Philippe Jund |
| Abstract Scope |
Thermoelectric (TE) materials have garnered attention since almost 60% of the fossil fuel energy is lost as heat. While machine learning can facilitate the accelerated discovery of efficient materials from the enormous configurational space of different compounds, the scarcity of experimental TE datasets limits the application of several conventional ML techniques. Moreover, their “black-box” nature fails to give structure-property relationships to chemists to develop design criteria for high performing materials. Herein, we propose using a novel symbolic regression-based ML technique, SISSO (Sure Independent Screening – Sparsifying Operator) to determine physically interpretable descriptors for predicting the TE figure-of-merit, zT, from a relatively small dataset of TE materials. Using SISSO for a promising class of TE materials (half-Heuslers) as an example, we demonstrate how, from a relatively small set of atomic features, a complex target property like zT can be predicted for novel compounds and confirmed by DFT calculations and experiments. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Modeling and Simulation, Characterization |