Can Public DFT Databases Guide Composition Selection for Metastable Magnets? A Validation of Energy-Above-Hull Screening Against Rapidly Solidified and Milled TbCu7-Type Sm-Fe-M Alloys

Public density-functional-theory (DFT) databases are increasingly used to screen compositions for permanent magnets. Yet it remains unclear how reliably their stability predictions apply to metastable phases produced by non-equilibrium processing. Here, we compare six public resources—Materials Project, AFLOW, JARVIS-DFT, NOMAD/Alexandria, OQMD, and GNoME—against published phase observations for melt-spun and mechanically milled Sm–Fe and Sm–Fe–M alloys (M = Co, Ti, V, Zr, Nb, Cu, Si, Ga, Al, Mo, Cr, Mn, or Hf). The database snapshot contains 628 entries; 30 of 38 curated literature observations can be scored across 11 alloy systems. Because none of the databases explicitly represents disordered TbCu7, we construct a stability proxy from the lower convex envelope of available ordered CaCu5, 2:17, 3:29, and ThMn12 relatives, measured against each database’s own convex hull. We also estimate the ideal configurational-entropy contribution. Materials Project, OQMD, and Alexandria place no binary Sm–Fe compound on their respective hulls. They assign closely spaced energies to equilibrium Sm2Fe17 (19.4 - 22.9 meV/atom), metastable TbCu7-type SmFe9 (proxy: 15.6–19.3 meV/atom), and SmFe12 (18.6–23.7 meV/atom), which is not obtained as a bulk ThMn12 phase by the melt-spinning routes considered. Their maximum separation is only 4.4 meV/atom, and the proxy for metastable SmFe9 is ranked below equilibrium Sm2Fe17. Consequently, their high hit rates at a 50 meV/atom threshold (20–24 hits, with at most one miss) do not establish an ability to distinguish phase-forming outcomes. JARVIS-DFT gives the expected ordering but places the SmFe9 proxy at 89.4 meV/atom, missing 19 of 20 covered positive observations. AFLOW and GNoME provide no comparable matches, and the screened resources contain no explicit disordered TbCu7, ω, or Sm5Fe17 entries. Interpolated Fe moments for SmFe9 differ from a measured value extrapolated to 0 K (1.88–1.96 μB/Fe) by +16% for Alexandria and −22% for JARVIS-DFT. These results show that energy above hull, as currently exposed by the databases examined, is insufficient by itself to select metastable Sm–Fe–M phases. We recommend checking whether a database reproduces known equilibrium phases before applying formability thresholds and provide the underlying proxy table, validation matrix, and reproducible pipeline.