UC San Diego

Micromagnetic simulation supports research on magnetic materials and devices. New applications require magnetization to evolve with moving or morphing geometry, varying material properties, and induced currents. Existing solvers address only parts of this problem, limiting detailed analysis. The main obstacles are long-range magnetostatic interactions, the airbox required by eddy-current methods, and the cost of updating both operators as the system evolves.These requirements call for a portable, memory-efficient compute foundation. SzCompute provides common memory containers and numerical operations for multicore CPUs, NVIDIA and AMD GPUs, and Apple Metal. Its provider model keeps the simulation state on the selected device and uses unified memory when the hardware supports it. DynaMag builds its finite-element solvers on this foundation. It retains fixed topology and degree-of-freedom identity while coordinates, material frames, and properties change. Matrix-free local kernels and reusable long-range plans then avoid large geometry-dependent matrices. This design makes accelerator portability part of the solver rather than a separate implementation for each device. On this foundation, the hybrid superposition integral–Poisson method addresses the magnetostatic field. It evaluates the free-space integral only on a truncation surface, after which a sparse Poisson solve recovers the interior field. B-AIM accelerates the nonlocal interaction, while matched local corrections retain finite-element accuracy. The point-to-point and finite-element corrections are evaluated as needed instead of being stored as matrices. The method therefore remains practical when the mesh changes and keeps the memory cost low. The eddy-current field is treated by coupling finite elements in the conductor to boundary elements on its surface. Nédélec elements represent the conducting region, while the boundary operator represents the unbounded exterior. This construction removes the airbox, reduces the number of unknowns, and avoids exterior remeshing. The same B-AIM infrastructure applies the boundary operator without forming a dense matrix and supports both free-space and periodic Green functions. B-AIM thus provides a common long-range engine for the magnetostatic and eddy-current solvers. Validation covers mesh motion, morphing, dynamic materials, magnetostatics, conduct- ing transients, and magnetization–eddy-current coupling. These tests define a low-memory path toward fully dynamic finite-element micromagnetics on heterogeneous hardware.