Machine learning based classification of vector field configurations

Abstract

Magnetic materials at the nanoscale are important for science and technology. A key aspect for their research and advancement is the understanding of the emerging magnetization vector field configurations within samples and devices. A systematic parameter space exploration—varying for example material parameters, temperature, or sample geometry—leads to the creation of many thousands of field configurations that need to be sighted and classified. This task is usually carried out manually, for example by looking at a visual representation of the field configurations. We report that it is possible to automate this process using an unsupervised machine learning algorithm, greatly reducing the human effort. We use a combination of convolutional auto-encoder and density-based spatial clustering of applications with noise (DBSCAN) algorithm. To evaluate the method, we create the magnetic phase diagram of a FeGe disc as a function of changing external magnetic field using computer simulation to generate the configurations. We find that the classification algorithm is accurate, fast, requires little human intervention, and compares well against the published results in the literature on the same material geometry and range of external fields. Our study shows that machine learning can be a powerful tool in the research of magnetic materials by automating the classification of magnetization field configurations.