Efficient Magnetic Hamiltonian Extraction from Magnetometry Using Machine Learning.

A magnetic material's spin texture arises from multiple competing interactions in the spin Hamiltonian, each of which favors a different spin configuration. Isolating and characterizing the strength of a single term in the Hamiltonian is therefore nontrivial. Of particular interest is the Dzyaloshinskii-Moriya interaction (DMI), which can stabilize topological spin textures such as magnetic skyrmions. Experimental techniques for the quantitative determination of DMI are often time consuming and require specialized instrumentation. Here we show that a convolutional neural network (CNN) can extract several Hamiltonian parameters from data obtained via magnetometry, the measurement of ensemble-averaged spins in magnetic samples. These CNNs were trained on HoyaFORCs 2.0, a database of simulated magnetometry measurements created for this work using fast algorithms for high-throughput hysteresis simulation. We further show that the predicted Hamiltonian can reproduce the magnetometry data through simulations and that their prediction uncertainty can be estimated by other CNNs that are trained on their errors. This process of estimating parameters from magnetometry measurements allows for an accelerated workflow in the exploration of topological spin textures for next-generation information processing.