AI-Assisted Identification of Magnetic Orders and Skyrmions
Abstract
Exotic magnetic orders in two-dimensional (2D) materials are attracting huge interest for energy-efficient spintronic applications, yet realizing robust high-temperature van der Waals antiferromagnets and topological magnetic states remains challenging. In this work, we develop a machine-learning framework for identifying magnetic orders and predicting magnetization using structural, compositional, and electronic information derived from the Materials Project. Fixed-length descriptors are constructed for two complementary tasks: ferromagnetic (FM) versus antiferromagnetic (AFM) classification and quantitative magnetization prediction. Magnetic order is classified using a LightGBM model trained on structure-derived descriptors without explicitly including magnetic descriptors. Five-fold cross-validation grouped by chemical system is used to reduce chemical leakage, together with hyperparameter and classification-threshold optimization. On an isolated test set, the classifier achieves a balanced accuracy of 93.6% in the testing set. Magnetization is predicted using a residual multilayer perceptron with compositional, structural, electronic, and task-specific initial-state descriptors. A logarithmic target transformation and robust weighted loss are used to account for the broad magnetization distribution, while three independently trained models are combined into a final ensemble. The model achieves a mean absolute error of 0.686 μB/formula unit. For selected magnetic candidates, simulated magnetic imaging and phase-reconstruction analysis are further used to investigate magnetization textures and skyrmion-like features through normalized magnetization profiles and topological charge. This framework provides an efficient approach for screening 2D magnetic materials and prioritizing candidates for antiferromagnetic and topological spintronic applications.
Related
- Derivation of the Thiele equation for magnetic skyrmions using the Mori-Zwanzig formalism
- Uniqueness and radial symmetry of magnetic skyrmions
- Spin-Orbit-Torque-Driven Recurrent Motion of Synthetic Antiferromagnetic Skyrmions for Magnetic Clocking
- Observation of Magnetic-Anisotropy Crossover and High-Temperature Skyrmions in the Dirac Magnet Fe3Ge with a Distorted Kagome Lattice
- The Agentic AI Kerr Lab: LLM Agents for Kerr Microscopy Experiments with Magnetic Skyrmions
BibTeX
@misc{yang2026aiassisted,
title = {{AI-Assisted Identification of Magnetic Orders and Skyrmions}},
author = {Haowen Yang and Sophia Huerta and Yingying Wu},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.35566},
eprint = {2609.35566},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.35566},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - AI-Assisted Identification of Magnetic Orders and Skyrmions
AU - Haowen Yang
AU - Sophia Huerta
AU - Yingying Wu
PY - 2026
DA - 2026/09/28
PB - arXiv
DO - 10.48550/arxiv.2609.35566
UR - https://arxiv.org/abs/2609.35566
AB - Exotic magnetic orders in two-dimensional (2D) materials are attracting huge interest for energy-efficient spintronic applications, yet realizing robust high-temperature van der Waals antiferromagnets and topological magnetic states remains challenging. In this work, we develop a machine-learning framework for identifying magnetic orders and predicting magnetization using structural, compositional, and electronic information derived from the Materials Project. Fixed-length descriptors are constructed for two complementary tasks: ferromagnetic (FM) versus antiferromagnetic (AFM) classification and quantitative magnetization prediction. Magnetic order is classified using a LightGBM model trained on structure-derived descriptors without explicitly including magnetic descriptors. Five-fold cross-validation grouped by chemical system is used to reduce chemical leakage, together with hyperparameter and classification-threshold optimization. On an isolated test set, the classifier achieves a balanced accuracy of 93.6% in the testing set. Magnetization is predicted using a residual multilayer perceptron with compositional, structural, electronic, and task-specific initial-state descriptors. A logarithmic target transformation and robust weighted loss are used to account for the broad magnetization distribution, while three independently trained models are combined into a final ensemble. The model achieves a mean absolute error of 0.686 μB/formula unit. For selected magnetic candidates, simulated magnetic imaging and phase-reconstruction analysis are further used to investigate magnetization textures and skyrmion-like features through normalized magnetization profiles and topological charge. This framework provides an efficient approach for screening 2D magnetic materials and prioritizing candidates for antiferromagnetic and topological spintronic applications.
ER -