Automated micromagnetic numerical simulations incorporating machine learning-based Fourier Neural Operators for efficient demagnetization calculations

Micromagnetic dynamics are fundamental to the development of advanced memory systems and computational technologies. Magnex, a solver for magnetic materials that solve the Landau-Lifshitz-Gilbert (LLG) equations, including exchange, anisotropy, demagnetization, and Dzyaloshinskii-Moriya interaction (DMI) coupling, was adapted to be a hybrid model in which the demagnetization numerical calculation was changed to be resolved using a Fourier Neural Operator (FNO), thus having integration of machine learning methodologies in a numerical solver to complement the simulation of micromagnetic dynamics. The FNO was trained on a database composed of a diverse applied field in the x and y directions with different magnitudes. For this, a Utility To Execute Pipeline (UTEP) was developed to efficiently organize and extract the outputs from simulations that are being systematically sent to the high-performance supercomputer Perlmutter. The accuracy and efficiency of this machine learning-numerical hybrid solver were validated using Micromagnetic Standard Problem #4 defined by the National Institute of Standards and Technology (NIST). The new dataset generated using this pipeline shows significant potential to support numerical simulations with machine-learning techniques.