Dynamics and relaxation

Nmag integrates the Landau-Lifshitz-Gilbert dynamics of the nodal magnetization. The default backend is SciPy DOP853 with relative and absolute tolerances of 1e-6 and a configured maximum step of 1 ps.

For meshes with exchange coupling, Nmag may derive a smaller effective maximum step to resolve the fastest lumped-FEM exchange mode. Inspect both the configured and effective values:

print(simulation.integrator_config)
print(simulation.effective_integrator_max_step)

Advance to a physical time

reached = simulation.advance_time(nmag.SI(10e-12, "s"))
print(reached, simulation.last_integrator_stats)

The target cannot precede the current stage time. max_it can cap accepted steps for interactive work, and exact_tstop controls whether the final state is reconstructed at the requested time.

Relax a state

simulation.relax()
print(simulation.time)
print(simulation.step)

The default convergence threshold is one degree per nanosecond. It is checked every five accepted integration steps and must be satisfied twice consecutively. Because that cadence counts accepted steps, a different integrator can confirm the same physical threshold at a different simulated time. Compare spatial stop states as well as averages when validating a workflow.

Tune a simulation before starting integration:

simulation.set_params(
    stopping_dm_dt=nmag.SI(0.5e9 * 3.141592653589793 / 180.0, "1/s"),
    ts_rel_tol=1e-7,
    ts_abs_tol=1e-7,
    ts_max_step=nmag.SI(0.5e-12, "s"),
)

Tighter tolerances do not compensate for an under-resolved mesh. Perform mesh, step, and tolerance convergence checks separately.

Scheduled saves

The default relaxation saves averages and fields at the end of the stage. A custom schedule uses when specifications:

from when import at, every

simulation.relax(
    save=[
        ("averages", every("step", 10)),
        ("fields", at("stage_end")),
        ("restart", at("stage_end")),
    ]
)

The public abbreviations normalize to save_averages, save_fields, and save_restart. Use simulated-time schedules when comparisons must be independent of adaptive step counts.

Experimental Diffsol backend

After building the Rust accelerator:

config = nmag.NmagConfig(integrator_backend="diffsol", accelerator="rust")
simulation = nmag.Simulation(config=config)

Diffsol is an opt-in dense implicit relaxation backend. It supports the default relaxation schedule for dense isotropic systems, but not advance_time, custom save/convergence schedules, anisotropy, or low-memory mode. SciPy remains the general production backend.