Multi-target machine learning and deep learning framework for predicting structural, energetic, and magnetic properties of Ternary Heusler alloys using site-resolved descriptors

Ternary Heusler alloys are promising for spintronic and magnetic applications, but their discovery via density functional theory (DFT) is computationally expensive. We develop a multi-target machine learning framework to predict structural, energetic, and magnetic properties from 839 DFT entries covering 420 compositions in both full-Heusler ( L 2 1 ) and inverse-Heusler (XA) prototypes. Descriptors are grouped by provenance, and models are evaluated using leave-composition-out cross-validation with a locked hold-out to prevent data leakage. Using only composition-level descriptors, the lattice constant is predicted with R 2 = 0.873 (MAE = 0.080 Å) and formation energy with R 2 = 0.812 (MAE = 0.079 eV/atom). Magnetic properties are more challenging: spin polarization reaches R 2 = 0.281 (MAE = 21.6 percentage points) and total magnetic moment R 2 = 0.353 (MAE = 0.94 μ B ), outperforming the Slater–Pauling (SP) rule. Including site-resolved sub-moments significantly improves spin polarization prediction to R 2 = 0.592 (MAE = 14.4 percentage points), highlighting the importance of electronic-structure information. Calibrated prediction intervals achieve near-nominal coverage for three of the four targets via split-conformal methods. Prototype preference prediction from composition attains 0.893 accuracy (ROC-AUC 0.958). Overall, the framework enables leakage-free, uncertainty-aware screening of Heusler alloys while clarifying the limits of composition-based prediction.