CQ-08: ML-based Magnetization Field Classification.

Magnetic materials at the nanoscale have attracted considerable interest both from academia and industry. A key aspect is the thorough understanding of the emerging magnetization field configurations, and their behaviour within samples and devices. However, an in-depth study such as plotting a magnetic phase diagram [1, 2] as a function of geometry, material parameter or applied field, can lead to creation of a large amount of data. Its evaluation typically requires considerable human effort and time. We report that it is possible to automate this process, to a certain degree, using an unsupervised machine learning (ML) algorithm. In this study, we use Agglomerative Clustering in magnetic phase diagram discovery of a FeGe disc with changing external magnetic field. First, we obtain the equilibrium magnetisation configurations by changing the field values from 0 to 1.2 T using micromagnetic simulations. Second, we obtain a feature space by passing the simulation data through the pre-trained layers of VGG-16 [3]. As the final step, we perform the Agglomerative Clustering on the feature space to sort the configurations in different classes. We find that the algorithm performs well when compared to the published results [1] on the same material geometry and range of external fields (Fig. 1). Although, the different classes obtained through the process do show some mixing of the equilibrium magnetisation configurations (Fig. 2). Our study shows that ML can be used to aid researchers in evaluating large amounts of magnetisation data. In our case, the algorithm clustered a thousand simulation results into individual classes without any training or human supervision, greatly reducing human effort for the task.References: [1] M. Beg et al., Sci. Rep. 5, no. 1 (2015) [2] S. Pathak et al., Phys. Rev. B. 103, no. 10 (2021) [3] K. Simonyan et al., ICLR, (2015)