ML rapidly screens more than 100,000 Heusler alloys to identify 80 thermoelectric candidates

1/ AI application method: spin-resolved d-DOS prediction with an atomistic line graph neural network
2/ Applied task: fast screening of thermoelectric candidates by predicting the electronic structure of 106,235 Heusler structures
3/ Reported result: the predicted DOS was validated by comparison with DFT DOS, 80 computational candidates were confirmed, and 61 of them were organized as unreported experimental candidates
4/ The researchers applied an atomistic line graph neural network that predicts spin-resolved d-DOS to Heusler alloy data, and rapidly screened thermoelectric material candidates across 106,235 structures in the DXMag database.
5/ The predicted density of states was validated by comparison with DFT calculations, and this information was combined with first-principles transport calculations to narrow the candidates.
6/ In the end, 80 computational candidates were organized, and 61 of them were presented as new candidates with no previously reported experimental data.
7/ It was also reported that the DOS-based Seebeck metric is effective for prioritizing top candidates.
8/ From a materials perspective, this result shows that linking electronic-structure prediction and transport metrics across a large composition and structure space can improve screening efficiency.
9/ However, the reported achievement is a computational screening result, and actual thermoelectric performance may vary depending on synthesis, phase stability, microstructure, doping capability, and measurement conditions. [1]