CuPyMag: GPU-Accelerated Finite-Element Micromagnetics with Magnetostriction

Functional magnetic materials often contain defects and grain boundaries with curved geometries across multiple length scales, which contribute to non-trivial magnetoelastic interactions. Capturing these effects makes micromagnetic simulation particularly demanding. To overcome this challenge, we introduce our recently developed program: CuPyMag, a Python framework with a GPU-resident workflow. After a one-time CPU assembly with Numba JIT, all subsequent operations are tensorized using CuPy's BLAS-accelerated backend. This design ensures high GPU utilization and minimizes host-device communication. In addition, we use the Gauss-Seidel projection method for stable and efficient time integration, and an ellipsoid theorem for the far-field effect of the demagnetization field. As a result, a magnetoelastic coupled system with 3M nodes is completed in 3 hours on a single H200 GPU. This efficiency expands the scope of micromagnetic simulations towards realistic, large-scale material problems that can guide experiments.