High-throughput screening of Heusler alloys for thermoelectrics combining machine-learned densities of states and first-principles transport
Thermoelectrics turn waste heat into electricity, but finding new ones requires transport calculations too costly for whole databases. Searches therefore choose what to calculate from the electronic density of states (DOS), increasingly predicted by machine learning, yet such choices are rarely tested. We test them on all 106,235 Heusler structures in the DXMag database, screening 51,257 on density-functional spectra and 54,978 on neural-network predictions, and judging both by 3,625 band-structure and transport calculations. A DOS-derived Seebeck indicator ranks most compounds passing every transport and stability test within the top 1% of their population, and ten of the eleven established half-Heusler thermoelectrics calculated pass. Predicted spectra lose most band gaps, unnoticed by spectral-similarity metrics, yet still rank passing compounds near the top. The screen certifies 80 computational candidates, 61 without reported measurements. Computed or predicted, the DOS can choose what to calculate; only a band structure shows whether a gap exists.