Quantifying Order in Magnetic Systems with Convolutional Neural Networks

Quantifying the order parameters in simulations of disordered magnetic systems to reliably represent magnetic states requires efficient interpretation of the often irregular magnetic patterns contained in large data sets. In helimagnetic materials with impurities, or at non-zero temperatures, the helical and skyrmion textures become distorted and embedded in noisy backgrounds with varying topological properties. Standard approaches based on quantifying the order parameters by averaging the spin states over a lattice, or computing pair correlation functions often become inconclusive. In this presentation, we discuss an approach based on applying convolutional neural networks (CNNs) to extract representative features of skyrmionic textures from simulated snapshots of complex magnetic backgrounds. Using large-scale micromagnetic simulations of a broad class of helimagnetic materials we show that CNNs are capable of not only accurately resolving the underlying magnetic textures, but also allow for regression to accurately predict classes of micromagnetic models representative of the observable magnetic textures. We then draw general conclusions about the uniqueness and invertibility of micromagnetic models of helimagnetic materials.