Modeling and prediction of magnetic material defect density thresholds using deep learning

Recent developments in deep learning techniques and data-driven methods have proven to be highly effective in modeling and predicting solutions to complex nonlinear systems. In the context of material science, predictions should be accompanied by limitations on the practical feasibility of solutions. We argue that such limitations can be found by combining deep learning techniques with suitable physical models. In the context of magnetic materials, we use the recently developed pseudospectral Landau-Lifshitz equation (PS-LL) [2] extended with the effect of material defects on the magnon dispersion relation. By adjusting a given ideal dispersion relation, we analytically model a defect size and density for distributed material defects in 1D using a random telegraph noise distribution [2]. Variational dispersion patterns for materials result in distinctive defect widths and density as a function of material parameters. We then apply convolutional neural networks to predict these error or defect thresholds based on simulated data. This approach promises to be a powerful tool for better understanding the possibilities and limitations on the discovery of functional magnetic materials.

[1] Kyle Rockwell, Joel Hirst, Thomas A. Ostler, and Ezio Iacocca, Phys. Rev. B 109, L180404 (2024)

[2] Stefan Machlup, J. Appl. Phys. 25, 341-343 (1953)