🔬 materials science

AI-Assisted Identification of Magnetic Orders and Skyrmions

This paper presents a machine-learning framework that utilizes structural, compositional, and electronic descriptors to accurately classify magnetic orders and predict magnetization in 2D materials, thereby enabling the efficient screening of candidates for spintronic applications and the identification of skyrmion-like features.

Original authors: Haowen Yang, Sophia Huerta, Yingying Wu

Published 2026-09-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Haowen Yang, Sophia Huerta, Yingying Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where the tiny magnets inside your computer could be flipped not by electricity, but by the gentle push of a magnetic field, consuming almost no energy in the process. This is the promise of spintronics, a field that seeks to use the intrinsic "spin" of electrons rather than their charge to store and process information. For this technology to work, scientists need materials that can host specific, exotic magnetic patterns. Among the most sought-after are skyrmions: tiny, swirling knots of magnetism that are incredibly stable and can be moved with very little energy. Finding these materials is like searching for a needle in a haystack, but the haystack is made of millions of different chemical combinations, and the needle is a specific arrangement of atoms that only reveals its secrets under the most extreme conditions. Traditionally, finding these materials has relied on slow, expensive computer simulations that calculate the behavior of every single electron, a process that is too time-consuming to screen the vast universe of possible materials.

A team of researchers at the University of Florida has now built a new kind of digital microscope to speed up this search. Instead of simulating every electron, they trained artificial intelligence to recognize the fingerprints of magnetic materials by looking at their chemical recipes and crystal shapes. The researchers developed two distinct AI tools to tackle the problem. The first acts as a sorter, quickly deciding whether a material is likely to be ferromagnetic, where all the tiny atomic magnets point in the same direction, or antiferromagnetic, where they point in opposite directions and cancel each other out. The second tool acts as a measurer, predicting exactly how strong the magnetism will be in a given material. By using these tools, the team was able to sift through a massive database of over 130,000 materials, identifying thousands of promising candidates that had previously been overlooked or misclassified.

The journey began with a challenge: the data available in existing scientific databases was often messy. The databases contained labels for magnetic states, but these labels were sometimes based on initial guesses made by computers rather than the final, stable state of the material. To fix this, the researchers wrote a new set of rules to reconstruct the true magnetic state of each material. They looked at the tiny magnetic moments of individual atoms within the crystal structure, checking if they were balanced or unbalanced, and whether they were aligned or opposed. They tested these rules against many different numerical thresholds to ensure that their conclusions were robust and not just a fluke of the calculation method. This careful cleaning process allowed them to build a reliable training set, separating materials into clear categories of non-magnetic, ferromagnetic, antiferromagnetic, and ferrimagnetic.

With a clean dataset in hand, the team turned to the first AI model: a classifier designed to distinguish between ferromagnetic and antiferromagnetic materials. Remarkably, they fed this model only information about the material's structure and composition—the types of atoms present, how they are arranged, and the geometry of the crystal lattice. They deliberately excluded any direct information about magnetism, forcing the AI to learn the hidden connections between a material's shape and its magnetic behavior. The AI learned that the way atoms are spaced and the specific elements they are made of contain subtle clues that determine whether the spins will align or cancel out. When tested on a set of materials it had never seen before, the model correctly identified the magnetic order of nearly 94 percent of the cases. It was particularly good at spotting the rare antiferromagnetic materials, which are difficult to find but essential for future energy-efficient devices.

The second task was more complex: predicting the exact strength of the magnetism. Unlike the first task, which was a simple yes-or-no question, this required the AI to predict a continuous number. The data for this was tricky because most materials have weak magnetism, while a few have extremely strong magnetism, creating a lopsided distribution. To handle this, the researchers transformed the data and used a specialized type of neural network, a deep learning model that can find complex patterns in high-dimensional data. This model was fed not just the structure and composition, but also details about the material's electronic bands and the initial settings used in the original computer simulations. The result was a highly accurate predictor that could estimate the magnetic strength of a material with an average error of less than one unit of magnetic moment per formula unit. This level of precision means the model can reliably rank materials, helping scientists decide which ones are worth studying in the lab.

To prove that their findings were not just numbers on a screen, the researchers took a step further and simulated the actual magnetic textures of a promising candidate material, a compound known as Fe3GaTe2. They used a technique that mimics how electrons pass through a thin slice of material to create an image of the magnetic field. By applying advanced image processing filters, they were able to reconstruct the magnetic phase and visualize the swirling patterns within the material. The simulation revealed a complex landscape of magnetic domains, including regions that resembled the skyrmions scientists are hunting for. This step confirmed that the materials identified by the AI were not just statistically likely to be magnetic, but were capable of hosting the intricate, topological structures needed for next-generation spintronic applications.

The work represents a significant shift in how new materials are discovered. By combining rigorous data cleaning with powerful machine learning, the researchers have created a framework that can screen vast chemical spaces in a fraction of the time it would take using traditional methods. The approach does not replace the need for detailed physical simulations or experiments, but it acts as a highly effective filter, narrowing down the millions of possibilities to a manageable list of high-priority candidates. The success of the models suggests that the secrets of magnetic behavior are encoded in the geometry and chemistry of materials, waiting to be decoded by algorithms that can see patterns invisible to the human eye. As the field moves forward, this method offers a clear path toward discovering the robust, room-temperature magnetic materials that will power the efficient, high-speed electronics of the future.

Technical Summary: AI-Assisted Identification of Magnetic Orders and Skyrmions

Problem Statement

The discovery of robust, high-temperature two-dimensional (2D) van der Waals antiferromagnets (AFM) and topological magnetic states (such as skyrmions) is critical for energy-efficient spintronics, neuromorphic computing, and quantum technologies. However, identifying these materials remains challenging due to the computational cost of traditional density functional theory (DFT) screening and the difficulty of distinguishing complex magnetic orders (e.g., AFM vs. ferromagnetic (FM) vs. ferrimagnetic (FiM)) based solely on structural data. Furthermore, existing databases often rely on categorical magnetic labels that may not accurately reflect the true magnetic ground state or the balance of magnetic sublattices. This work addresses the need for an efficient, machine-learning (ML)-driven framework to screen 2D magnetic materials for specific ordering types and quantitative magnetization values without performing exhaustive first-principles calculations for every candidate.

Methodology

1. Dataset Curation and Magnetic State Reconstruction

The authors utilized data from the Materials Project (271,325 initial entries) but did not use the database's categorical magnetic labels directly as ground truth. Instead, they reconstructed magnetic states from site-resolved DFT magnetic moments (magmom) to ensure physical accuracy.

Reconstruction Logic: The team defined active magnetic sites based on a moment threshold ( $m_{th} = 0.20 \, \mu_B$ ). They calculated the absolute moment ( $M_{abs}$ ), net moment ( $M_{net}$ ), and a normalized alignment parameter ( $A$ ) and sublattice balance ( $B$ ).
Labeling Criteria:
- AFM: Requires $A \le 0.05$ (high compensation) and $B \ge 0.90$ (balanced sublattices).
- FM: All active moments aligned in one direction.
- FiM: Oppositely oriented sublattices with significant imbalance.
- NM: No active sites.
Robustness Check: To avoid sensitivity to numerical cutoffs, the labeling was tested across 15 combinations of thresholds ( $m_{th}$ and $A_{max}$ ). A material retained its label only if the assignment was consistent in at least 12 of the 15 cases ( $R \ge 0.80$ ).
Data Recovery: For materials lacking final task-level moments, a targeted recovery process extracted site moments from material structures for known AFM and FiM candidates, significantly expanding the robust AFM dataset from 121 to 3,485 entries.
Final Dataset: The curated dataset contained 21,789 materials (18,304 FM, 3,485 AFM) for the classification task.

2. Magnetic Order Classification (FM vs. AFM)

Input Features: 136 fixed-length descriptors derived strictly from structural, compositional, and elemental properties (e.g., lattice constants, coordination numbers, composition entropy, elemental fractions). Crucially, no explicit magnetic descriptors (e.g., magnetic moments, spin polarization) were included in the input to prevent data leakage.
Model: A LightGBM gradient-boosted decision tree classifier.
Validation Strategy: Five-fold cross-validation grouped by chemical system to prevent chemical leakage (ensuring chemically similar materials do not appear in both training and test sets).
Optimization: Hyperparameters were tuned using Optuna. The decision threshold for classifying AFM was optimized to 0.318 to handle class imbalance.

3. Magnetization Value Prediction (Regression)

Input Features: A broader set of 464 raw descriptors (expanded to 1,264 after encoding) including composition, structure, electronic band structure (band gaps, DOS near Fermi level), and calculation metadata (initial magnetic moments, Hubbard-U, run type). Final magnetic moments were explicitly excluded.
Target: Total magnetization normalized per formula unit ( $\mu_B$ /f.u.), which exhibited a highly right-skewed distribution.
Preprocessing:
- Target Transformation: Applied a RobustLog transformation ( $y' = \log(1+y)$ ) to compress the dynamic range.
- Weighting: A magnitude-dependent weighting scheme was used to increase the contribution of high-magnetization samples during training.
Model: A residual multilayer perceptron (MLP) with three residual blocks, trained with a weighted Smooth-L1 loss.
Ensemble: Three models trained with different random seeds (42, 137, 2026) were averaged to form the final predictor.

4. Magnetic Image Processing and Skyrmion Identification

Simulation: Simulated Lorentz transmission electron microscopy (LTEM) images of $Fe_3GaTe_2$ containing skyrmion and stripe domains were generated using MuMax3.
Phase Reconstruction: The Single-Image Transport of Intensity Equation (SITIE) was used to reconstruct magnetic phase shifts from the LTEM images.
Optimization: Tikhonov regularization was applied to filter noise. The cutoff frequency ( $q_c$ ) was optimized to maximize the crest signal-to-noise ratio (SNR), identifying $q_c = 0.016 \, nm^{-1}$ as optimal.
Analysis: A lightweight CNN (TinyCNN) was proposed to map HSV color space from images to 3D vector fields for texture analysis.

Key Results

Classification Performance

On an isolated test set of 4,357 materials (chemically distinct from the training set):

Balanced Accuracy: 93.6%.
AFM Class Performance:
- Precision: 88.4%
- Recall: 89.5%
- F1 Score: 89.0%
- PR-AUC: 0.955
Observation: The model successfully distinguished AFM from FM configurations using only structural and compositional data, demonstrating that these properties encode sufficient information about magnetic exchange pathways.

Regression Performance

On the same isolated test set:

Mean Absolute Error (MAE): 0.686 $\mu_B$ /f.u.
Root-Mean-Square Error (RMSE): 1.769 $\mu_B$ /f.u.
Coefficient of Determination ( $R^2$ ): 0.905.
Error Distribution: Errors were lowest for low-magnetization materials (<0.5 $\mu_B$ /f.u., MAE $\approx$ 0.25) and increased for high-magnetization materials (>20 $\mu_B$ /f.u., MAE $\approx$ 3.7), reflecting the scarcity of high-moment samples in the training data.

Image Analysis

The SITIE analysis with optimized Tikhonov filtering successfully reconstructed magnetic phase textures from simulated LTEM images, revealing labyrinth-like and skyrmion-like features with improved spatial contrast and reduced noise.

Significance and Claims

The paper claims to provide an efficient, high-throughput framework for screening 2D magnetic materials that prioritizes candidates for antiferromagnetic and topological spintronic applications.

Data-Centric Rigor: By reconstructing magnetic labels from site-resolved moments rather than relying on database categories, the authors created a more physically robust dataset, specifically increasing the reliability of the minority AFM class.
Structure-Property Correlation: The classification results demonstrate that magnetic ordering (FM vs. AFM) can be predicted with high accuracy using only non-magnetic structural and compositional descriptors, suggesting that the "statistical signatures" of magnetic exchange are embedded in the crystal geometry and chemical composition.
Quantitative Prediction: The regression model offers a quantitative estimate of magnetization strength, enabling researchers to filter materials not just by order type but by magnetic magnitude.
Topological Identification: The integration of ML with image processing techniques (SITIE, Tikhonov filtering) provides a pathway to identify and reconstruct complex spin textures like skyrmions from simulated or experimental imaging data.

The authors position this work as a "first-stage screening framework" that narrows down the vast chemical space of 2D materials, reducing the computational resources required for subsequent, more expensive first-principles calculations or experimental synthesis. They explicitly note that the current models predict ground-state properties at 0 K and do not yet account for finite-temperature transitions (Curie or Néel temperatures), which they identify as a necessary future direction.