Applying Machine Learning Techniques to Skyrmion Phase Diagrams

F. A. GÓMEZ ALBARRACÍN; H. D. ROSALES

Conferencia; 33rd IUPAP Conference on Computational Physics; 2022

University of Texas at Austin

Skyrmions are swirling arrangements of spins (magnetic moments), with topological characteristics that make them potentially relevant for technological applications. In the magnetic model chosen in this work, there are three very clearly defined low temperature phases: spirals, skyrmions crystal andferromagnetic (or fully polarized), which can be characterized with the structure factor and order parameters such as the scalar chirality. Previous works have shown that neural networks can successfully classify these phases. Here, we aim to produce a complete phase diagram, including the more intricate and less defined intermediate phases (bimerons or elongated skyrmions, and skyrmion gas). In order to do this, we construct a convolutional neural network (CNN) which we train and validate for classification using only configurations obtained from high performance Monte Carlo simulations at the lowest simulated temperature, ensuring in this way that the structures are well defined. These configurations correspond only to three values of the parameters of the model(which implies three different skyrmion and spiral sizes). We emphasize that we do not use any other parameter, such as temperature for example, as input. We also include very high temperature disordered snapshots from the paramagnetic phase. Then, we test the CNN for other parameters (skyrmion sizes), that were not used in training or validation. Furthermore, we apply the resulting CNN model to the whole set of generated spin configurations, in the complete range of temperature and magnetic field, to construct a detailed phase diagram, with very good results. We also compare with other simpler machine learning methods, such as Support Vector Machine and Random Forest.