Pre-Refinement Prediction of Dopant Site Occupancy in M-Type Hexaferrites via Hierarchical and Uncertainty-Aware Machine-Learning

Dopant site occupancy bridges chemical composition, local crystal structure, and macroscopic functional responses in M-type hexaferrites. However, reliable occupancy hypotheses are difficult to establish before structural refinement. Here, we develop a pre-refinement machine-learning framework using 494 dopant-level records extracted from refined structures. This workflow combines Ba/Sr-vs.-Fe sublattice classification, fractional occupancy prediction over five inequivalent Fe sites, and reliability-aware assessment. The CatBoost classifier achieves 96% accuracy on external validation for sublattice assignment. For Fe-sublattice dopants, homogeneous occupancy-model screening and heterogeneous cooperative occupancy-model screening yield a macro mean absolute error as low as 0.0584, while preserving dominant-site identification and occupancy ranking. The results show that dopant site occupancy is jointly controlled by dopant chemistry, target-dopant amount, and co-doping environment rather than a universal site-preference rule. By integrating TabPFN baseline comparison, training-experience assessment, and model-consistency analysis, the framework identifies boundary and extrapolative samples that require priority refinement. Our predictions for W-doped and Sc-doped M-type hexaferrites successfully reproduce key experimental occupancy trends. This work offers a practical data-driven strategy to generate physically plausible occupancy hypotheses and assign refinement priorities for partially occupied oxide systems.