This study investigates neural network approaches for predicting local magnetic moments of iron atoms in Fe-Al disordered alloys, using a domain-specific dataset based on quantum-mechanical density functional theory (DFT) calculations. The data set includes $6880$ iron atoms and $3420$ aluminum atoms, divided into train, validation, and test sets, with model development and evaluation focusing exclusively on iron atoms. The average local magnetic moment of iron is measured as $2.013\,\mu_B$ with a standard deviation of $0.424\,\mu_B$.

Local environments are encoded using the Smooth Overlap of Atomic Positions (SOAP) descriptor from the dscribe library, with a cutoff radius capturing the first three nearest coordination shells around each atom. This yields feature representations invariant to rotation and translation, while focusing on relevant chemical and geometric correlations.

Our feedforward neural network, trained with SOAP descriptors, predicts iron atom magnetic moments with a mean absolute error (MAE) of $0.0436\,\mu_B$ on the test set. For comparison, the baseline model that assigns to each iron atom the mean magnetic moment observed for its specific first-neighbor configuration yields an MAE of $0.1205\,\mu_B$ . On the same task, CHGNet achieves an MAE of $0.2147\,\mu_B$ , demonstrating that our SOAP-based approach provides significantly higher accuracy in predicting local magnetic moments in Fe-Al alloys, while requiring a fraction of memory and computing time.

The results highlight the effectiveness of SOAP-based neural network models for fast and accurate prediction of electronic properties in metallic alloys, with implications for large-scale computational screening and the study of local magnetic phenomena.