Supported scope and limitations

Supported simulation model

Nmag for Python 3 supports 3D, first-order tetrahedral micromagnetic workflows with:

• legacy Nmesh HDF5 and modern Meshio-supported mesh input;
• single- and multi-region material assignment where material degrees of freedom are compatible at shared nodes;
• demagnetization, exchange, uniform applied field, uniaxial and cubic anisotropy, and custom polynomial anisotropy energy;
• pinning and Zhang-Li current-density spin-transfer torque;
• adaptive time advancement and relaxation with SciPy DOP853;
• NDT/HDF5 data, spatial field probing, and native restart checkpoints; and
• optional Rust kernels and resource-aware dense, hierarchical, or matrix-free demagnetization storage.

Important limitations

The following are not currently supported:

• thermal or stochastic dynamics;
• Slonczewski spin-transfer torque;
• periodic micromagnetic boundary conditions;
• custom phi_BEM/legacy HLib configuration;
• material-specific magnetization degrees of freedom where shared geometric nodes require conflicting material coefficients;
• local inter-material exchange coupling;
• spatially or time-varying applied fields;
• complete multi-stage legacy hysteresis compatibility;
• legacy restart-file import; and
• a distributed-memory MPI solver.

The experimental Diffsol backend has a narrower scope: dense isotropic relaxation with the default save and convergence schedule. It does not support anisotropy, low-memory mode, custom schedules, or public advance_time.

Numerical responsibility

Passing a solver test does not establish that a scientific model is resolved. For each new workflow:

1. verify material values and units;
2. check geometry, region tags, tetrahedron quality, and boundary resolution;
3. refine the mesh and compare relevant observables;
4. tighten tolerances and maximum time step independently;
5. inspect maximum neighboring spin angle and spatial fields; and
6. compare against an analytic result, trusted package, experiment, or a previously validated Nmag case.

Adaptive solvers need not take the same internal steps or stop at the same time to represent equivalent physical states. Compare field and magnetization arrays, energies, and problem-specific observables rather than requiring identical step histories.

Resource limits

There is no universal mesh-size limit. Memory and time depend strongly on the number of volume and boundary nodes, geometry, selected demagnetization storage, and requested dynamics. Dense boundary storage grows quadratically. Auto mode uses available-memory estimates, while low-memory mode avoids dense BEM storage at the cost of repeated computation.

Review last_bem_operator_stats, host memory, and elapsed setup/action time before scaling a study. Never disable a resource guard without estimating the resulting allocation.