Estimation of magnetic parameters from domain images with convolutional neural networks

Magnetic multilayer films are known to host a variety of novel magnetic configurations such as topological magnetic skyrmions with potential nano-electronic applications. However, characterizing these material systems can be time-consuming and expensive. Therefore, it is crucial to maximize the information extracted from results of experiments that are more accessible, such as domain images obtained from magnetic force microscopy. We show that deep convolutional neural networks are able to extract magnetic parameters such as the exchange interaction, Dzyaloshinskii-Moriya interaction, and uniaxial anisotropy from images of domain configurations. Experimentally realistic training and validation data were generated through micromagnetic simulations. The trained models were consistently able to reach R^2 values greater than 0.9 on validation data. By inspecting the intermediate feature maps of the neural network, we find that the network is able to learn features such as domain boundaries. Testing the models on actual experimental data yield values that were consistent with our knowledge of the material systems. Our work thus demonstrates the utility of developing machine models trained on simulation data as a means to accelerate the characterization of magnetic systems.