Detection of Nonuniformity in Parameters for Magnetic Domain Pattern Generation by Machine Learning

Naoya Mamada, Masaichiro Mizumaki, Ichiro Akai, Toru AonishiJournal of the Physical Society of Japan 2024 · DOI

TL;DR

This paper introduces a machine learning approach using convolutional neural networks (CNNs) to map the spatial distribution of physical parameters in magnetic thin films from their domain patterns. By leveraging large-scale CNNs and pretraining, they achieve improved accuracy in estimating these parameters, enabling the detection of non-uniformities crucial for material design and characterization.

What Problem Does This Paper Address?

Understanding and controlling the properties of magnetic thin films is vital for applications ranging from data storage to spintronic devices. These properties are influenced by various physical parameters (like noise coefficient D and exchange interaction J) that can vary spatially within the material. Traditionally, determining these parameters relies on empirical rules and trial-and-error, which is inefficient. While machine learning has shown promise in estimating these parameters from magnetic domain patterns, accurately mapping spatially varying parameters, especially from small subregions of the pattern, remains a challenge. This paper aims to address this by developing a robust CNN-based method to precisely map these spatially heterogeneous parameters.

Methods

The core of this work is the application of Convolutional Neural Networks (CNNs) for parameter estimation. The authors propose a "sliding window" approach: they train a CNN to estimate parameters (D and J) from small subregions of magnetic domain patterns. This trained model is then slid across the entire domain pattern, generating a spatial map of the estimated parameters.

To enhance accuracy, especially for small subregions, they employ two key strategies:

  • Large-Scale Models: They utilize large CNN architectures, specifically RegNetX-16GF, which are designed for complex natural image classification tasks. These models have a greater capacity to learn intricate patterns compared to smaller, custom-built CNNs.
  • Pretraining: They leverage transfer learning by pretraining the large CNNs on the ImageNet dataset (a large collection of natural images). This allows the model to learn general image features that are then fine-tuned for the specific task of analyzing magnetic domain patterns.

They developed two CNN models: a small-scale model (Model0-6) based on prior work and a large-scale model (reg) based on RegNetX-16GF. They also experimented with a pretrained version of the large-scale model (reg-pret).

Datasets

The study utilizes simulated magnetic domain pattern data generated through micromagnetic simulations based on the time-dependent Ginzburg-Landau equation.

  • Spatially Uniform Data: Used for training and initial model evaluation. This dataset consists of patterns generated with various combinations of the parameters D and J, where both parameters are uniform across the entire pattern.
  • Spatially Inhomogeneous Data: Used for evaluating the model's ability to detect parameter variations. This dataset features patterns where one parameter (D) varies spatially within the image, while the other (J) is fixed. The simulation space was discretized into 512x512 pixels, with a distinct "inner" region where D varied and an "outer" region where D was fixed.

Key Insights and Ideas

  • Large-scale CNNs are crucial for subregion analysis: The study highlights that smaller CNNs struggle to accurately estimate parameters from small subregions of magnetic domain patterns. Large-scale models, designed for complex image recognition, are significantly more effective.
  • Pretraining with natural images boosts performance: Leveraging ImageNet pretraining significantly improves the accuracy of parameter estimation, especially when training data is limited. This suggests that general visual features learned from natural images are transferable to understanding magnetic domain structures.
  • Estimating D is more challenging than J: The research consistently found higher estimation errors for the noise parameter D compared to the exchange interaction parameter J. This is attributed to the less pronounced visual changes in domain patterns as D varies, making it harder for the model to discern subtle differences.
  • Detecting spatial variations is feasible but requires resolution improvement: The proposed method can detect spatial variations in parameters, even with relatively small differences. However, achieving high-resolution mapping of these variations remains an area for improvement.

Key Results

  • Improved Accuracy with Large-Scale Models and Pretraining: The pretrained large-scale model (reg-pret) consistently outperformed the small-scale model (mod) and the non-pretrained large-scale model (reg) in estimating both D and J. For instance, with limited data (Ndata=4), reg-pret achieved mean absolute errors of 0.212 for D and 0.151 for J.
  • Significant Gains from Pretraining: Pretraining led to a notable improvement in accuracy, particularly for parameter D when data was scarce (Ndata=1), reducing the mean absolute error by 0.042.
  • Detection of Spatial Variations: The ROC analysis showed that the method could detect spatial variations in D with an AUC score of 0.798 even when the difference between inner and outer D was as small as 0.3.
  • Model Sensitivity: The large-scale models (reg and reg-pret) demonstrated better sensitivity to parameter changes compared to the small-scale model (mod), which showed insensitivity to D variations.

Limitations and Future Work

  • Resolution of Spatial Parameter Mapping: While the method can detect parameter changes, the current resolution for mapping these variations is limited. Future work should focus on increasing this resolution.
  • Outlier Detection: The sliding window approach, by using small patches, can lead to outliers in parameter estimation. Increasing the field of view of the input image while maintaining locality is suggested.
  • Interdependence of Parameters: The estimation of one parameter (J) can be influenced by variations in another (D). Improving the stability of parameter estimation to be independent of other parameters is a future direction.
  • Real-World Data Application: The current study relies on simulated data. Applying the method to real material data and more realistic polycrystalline simulations is a crucial next step.
  • Interpretability: Addressing the interpretability issues common to CNNs is also highlighted as a necessary improvement.
  • Fully Convolutional Networks: The authors suggest exploring fully convolutional networks (FCNs) for pixel-wise regression or segmentation, which could potentially detect parameter changes at higher resolutions and reduce outliers.
  • Bayesian Inference: For improving estimation around boundaries and providing confidence estimates, Bayesian inference is proposed.

Summary updated

Abstract

We estimate the spatial distribution of heterogeneous physical parameters involved in the formation of magnetic domain patterns of polycrystalline thin films using convolutional neural networks. We propose a method of obtaining a spatial map of physical parameters by estimating the parameters from patterns within a small subregion window of the full magnetic domain and subsequently shifting this window. To improve the accuracy of parameter estimation in such subregions, we employ large-scale models for natural image classification and exploit the benefits of pretraining. Using a model with high estimation accuracy on such subregions, we conduct inference on simulation data featuring spatially varying parameters and demonstrate the capability of this model to detect such parameter variations.
BibTeX
@article{mamada2024detection,
  author = {Naoya Mamada and Masaichiro Mizumaki and Ichiro Akai and Toru Aonishi},
  title = {{Detection of Nonuniformity in Parameters for Magnetic Domain Pattern Generation by Machine Learning}},
  journal = {Journal of the Physical Society of Japan},
  year = {2024},
  doi = {10.7566/jpsj.93.054706},
}
Figure