Dresden 2026 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MA: Fachverband Magnetismus

MA 34: Computational Magnetism I

MA 34.5: Talk

Wednesday, March 11, 2026, 16:00–16:15, HSZ/0004

Prediction of Iron Atom Magnetic Moments in Fe-Al Alloys via SOAP-based Neural Network Models — •Vojtěch Ráliš^1,2, Martin Friák¹, Jan Fikar¹, and Aleš Horák² — ¹Institute of Physics of Materials, v.v.i., Czech Academy of Sciences, Brno, Czech Republic — ²Faculty of Informatics, Masaryk University, Brno, Czech Republic

We investigate neural network approaches for predicting local magnetic moments of iron atoms in Fe-Al disordered alloys using a dataset of 227 unit cells (27–216 atoms) with Al concentrations of 0–60%, totaling 6880 Fe atoms (mean: 2.013 µ_B, std: 0.424 µ_B) from DFT calculations.

Local environments are encoded using SOAP descriptors with a cutoff capturing the first three coordination shells, yielding rotation- and translation-invariant features. Our feedforward neural network achieves MAE = 0.0436 µ_B on the test set, significantly outperforming a first-neighbor baseline (MAE = 0.121 µ_B) and CHGNet (MAE = 0.215 µ_B), while requiring far less memory and compute time.

These results demonstrate the effectiveness of SOAP-based neural networks for fast, accurate prediction of electronic properties in metallic alloys, with implications for large-scale computational screening and local magnetic phenomena studies.

Keywords: neural network; SOAP descriptor; DFT; local magnetic moment; local atomic environment