Regularized Micromagnetic Model

• Regularized micromagnetic models are advanced formulations that relax fixed-length assumptions to resolve singularities while preserving fundamental balance laws.
• They utilize methods such as higher-order exchange, measure-valued relaxation, and variable-length order parameters to overcome the limitations of classical micromagnetics.
• These techniques improve numerical stability, capture fine-scale dynamics, and enable effective multiscale modeling across varying thermal regimes.

A regularized micromagnetic model is a micromagnetic formulation in which the classical assumptions of fixed magnetization length, strictly local continuum closure, or fully resolved deterministic fields are modified to remove a specific pathology while retaining the micromagnetic balance laws relevant to the regime of interest. In the arXiv literature, regularization appears in several technically distinct forms: mesoscopic relaxation of nonconvex Brown micromagnetics by Young measures and nonlocal penalization, higher-order exchange terms that add a bi-Laplacian contribution to stochastic Landau–Lifshitz–Gilbert dynamics, variable-length order-parameter theories for Bloch points, exchange-damping regularizations of Landau–Lifshitz–Bloch/Baryakhtar type, and multiscale atomistic/continuum frameworks that treat the near- and above-$T_c$ region atomistically while retaining conventional micromagnetics in cooler regions (Benešová et al., 2012, Chugreeva et al., 2017, Kuchkin et al., 27 Aug 2025, Soenjaya, 2024, Gija et al., 4 Feb 2025).

1. Classical micromagnetics and the sources of nonregularity

Classical micromagnetics starts from a magnetization field subject to exchange, anisotropy, magnetostatic, and Zeeman interactions. In the Brown static setting used as the point of departure in the thermomagnetic mesoscopic model, the Gibbs energy contains the anisotropy density $\psi(m,\theta)$, the demagnetizing term $\frac12 m\cdot \nabla u_m$, the exchange penalty $\frac{\varepsilon}{2}|\nabla m|^2$, and the Zeeman coupling $-h\cdot m$, with magnetostatics constrained by $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$ on $\mathbb R^d$ (Benešová et al., 2012). In stochastic and dynamical micromagnetics, the standard unit-vector formulation likewise fixes the state on $S^2$, so the exchange field is built from the Dirichlet energy and the dynamics preserve $|m|=1$ (Chugreeva et al., 2017).

The need for regularization arises from several distinct failure modes. In relaxed Brown micromagnetics, minimizing sequences may develop arbitrarily fine domain patterns when the exchange parameter is small, and classical minimizers may fail to exist when exchange is negligible; the ferro/paramagnetic transition additionally requires thermodynamic coupling to temperature that standard static micromagnetics does not provide (Benešová et al., 2012). In Bloch-point dynamics, the hedgehog texture produces an exchange field diverging like $1/r^2$ at the core, so the fixed-length continuum theory becomes singular (Kuchkin et al., 27 Aug 2025). Near and above the Curie temperature, the fixed-norm Landau–Lifshitz model is conceptually invalid because spontaneous magnetization is no longer constant and above $\psi(m,\theta)$0 there is no spontaneous ferromagnetic order at all (Kiefe et al., 28 May 2026). In ferrimagnets, effective one-medium reductions can become problematic near compensation points, whereas a two-sublattice formulation keeps the sublattice equations well defined without introducing divergent effective parameters (Martínez et al., 2019).

These pathologies are not all of the same type. Some are variational nonattainment problems, some are singular-core problems, some are constitutive breakdowns, and some are numerical stiffness or closure deficiencies. The literature therefore uses “regularized micromagnetic model” in more than one sense. This suggests that the term is best read as a family resemblance concept rather than a single canonical construction.

2. Mesoscopic relaxation by Young measures

The most explicit variational regularization in the supplied literature is the thermodynamically consistent mesoscopic model for the ferro/paramagnetic transition. Its first regularization step is a measure-valued relaxation of nonconvex micromagnetic energy: the oscillatory magnetization field is replaced by a Young measure $\psi(m,\theta)$1, with macroscopic magnetization given by the first moment

$\psi(m,\theta)$2

The relaxed static problem is then posed over $\psi(m,\theta)$3, so unresolved microstructure is stored in $\psi(m,\theta)$4 while $\psi(m,\theta)$5 is its barycenter (Benešová et al., 2012).

The second regularization step is the introduction of an internal variable

$\psi(m,\theta)$6

followed by a deliberate relaxation of the exact moment constraint. Instead of imposing $\psi(m,\theta)$7 pointwise, the model penalizes the mismatch through the nonlocal $\psi(m,\theta)$8-term

$\psi(m,\theta)$9

The authors interpret $\frac12 m\cdot \nabla u_m$0 as being “in a position of a phase field.” Dissipation is encoded by

$\frac12 m\cdot \nabla u_m$1

so the model combines rate-independent hysteretic losses with viscous dissipation (Benešová et al., 2012).

The magnetic part remains quasistatic, whereas $\frac12 m\cdot \nabla u_m$2 and temperature evolve. After the enthalpy transformation $\frac12 m\cdot \nabla u_m$3, the coupled system consists of pointwise-in-time minimization of the regularized Gibbs functional in $\frac12 m\cdot \nabla u_m$4, a variational flow rule for $\frac12 m\cdot \nabla u_m$5, and a heat equation for $\frac12 m\cdot \nabla u_m$6. The weak solution concept mirrors this fast/slow split, and existence of at least one weak solution is proved by a semi-implicit backward Euler discretization, coercive time-discrete regularization, compactness, and lower semicontinuity arguments (Benešová et al., 2012).

Within this framework, “regularization” is explicitly threefold: mesoscopic relaxation by Young measures, $\frac12 m\cdot \nabla u_m$7-penalization of the moment constraint, and incremental coercive regularization in the time-discrete problem. The main limitation stated in the source is equally specific: for fixed finite $\frac12 m\cdot \nabla u_m$8, $\frac12 m\cdot \nabla u_m$9 is only approximately equal to the true moments of $\frac{\varepsilon}{2}|\nabla m|^2$0, and the full thermally coupled exact-constraint limit is only heuristically justified there rather than completely proved (Benešová et al., 2012).

3. Higher-order exchange and exchange-damping regularizations

A second major strand replaces the standard Dirichlet exchange energy by a higher-order functional. In the regularized stochastic Landau–Lifshitz–Gilbert equation, the exchange energy is

$\frac{\varepsilon}{2}|\nabla m|^2$1

so the effective field becomes $\frac{\varepsilon}{2}|\nabla m|^2$2. The corresponding Stratonovich SPDE preserves the geometric LLG structure while adding a bi-Laplacian contribution. This regularization is physically motivated by advanced exchange models connected with nanoscale topological solitons and is mathematically decisive: unlike the classical stochastic LLG based only on Dirichlet exchange, the regularized equation is globally solvable in the stochastically strong sense in dimensions two and three, pathwise unique, and yields trajectories in $\frac{\varepsilon}{2}|\nabla m|^2$3. The consequence stressed by the paper is almost-sure preservation of the topology of the initial data (Chugreeva et al., 2017).

The same higher-order logic reappears in finite-temperature continuum models through exchange damping. In the thermodynamically self-consistent LLBar formulation, the evolution law

$\frac{\varepsilon}{2}|\nabla m|^2$4

separates local relativistic relaxation from nonlocal exchange relaxation. After expansion and normalization, the resulting model contains a nonlocal damping contribution $\frac{\varepsilon}{2}|\nabla m|^2$5, longitudinal relaxation, and explicit coupling to electron and phonon temperatures, thereby extending standard micromagnetics beyond fixed-length LL/LLG and beyond single-bath damping closures (Dvornik et al., 2014).

A numerically and analytically unified formulation is developed for LLBar and the regularised LLBloch equation in mixed form,

$\frac{\varepsilon}{2}|\nabla m|^2$6

Here $\frac{\varepsilon}{2}|\nabla m|^2$7 is the regularizing exchange-damping parameter. Mixed finite element schemes based on semi-implicit Euler and semi-implicit Crank–Nicolson are proved unconditionally energy-stable and optimal in $\frac{\varepsilon}{2}|\nabla m|^2$8, $\frac{\varepsilon}{2}|\nabla m|^2$9, and $-h\cdot m$0 norms; the paper also proves convergence of LLBar to LLBloch as $-h\cdot m$1 with rate $-h\cdot m$2 for strong solutions (Soenjaya, 2024).

Taken together, these papers define a precise continuum meaning of regularization: one adds higher-order spatial control, either directly in the energy or through $-h\cdot m$3, in order to recover stronger well-posedness, preserve dissipativity at the discrete level, and control regimes where the unregularized model is analytically or numerically fragile.

4. Variable-length order parameters and Bloch-point dynamics

Bloch points motivate a different regularization strategy: replace the unit-vector order parameter by a bounded variable-length field. The regularized theory for Bloch points introduces a four-component field

$-h\cdot m$4

with physical magnetization $-h\cdot m$5 satisfying $-h\cdot m$6 and

$-h\cdot m$7

The regularized exchange density is

$-h\cdot m$8

which creates a characteristic Bloch-point core size $-h\cdot m$9 (Kuchkin et al., 27 Aug 2025).

This construction removes the singular effective field by allowing the magnetization magnitude to collapse continuously to zero at the Bloch-point core while remaining bounded above by saturation. The associated regularized LLG equation is formulated on $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$0 rather than $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$1,

$$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$2

with $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$3, $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$4, and $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$5. In the regime studied in the paper, the $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$6-term is set to zero (Kuchkin et al., 27 Aug 2025).

The practical consequence is not merely formal regularity. In BP-hosting textures such as nanowire domain walls, chiral bobbers, and dipolar strings, the $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$7 model removes mesh-dependent Bloch-point pinning, suppresses spurious oscillatory dynamics, eliminates artificial threshold current or field effects, and restores agreement with a generalized Thiele equation for steady motion. The model reduces to standard micromagnetics when $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$8, so the regularization is localized to the singular-core regime rather than being a wholesale replacement of ordinary micromagnetic dynamics (Kuchkin et al., 27 Aug 2025).

This bounded-amplitude strategy differs from Landau–Lifshitz–Bloch and LLBar regularizations. The paper explicitly argues that the $$\operatorname{div}(\mu_0\nabla u_m-\chi_\Omega m)=0$$9 constraint enforces the physically natural inequality $\mathbb R^d$0, whereas Landau-type amplitude theories do not inherently encode that upper bound and do not supply the same Thiele-type reduction for Bloch-point-carrying textures (Kuchkin et al., 27 Aug 2025).

5. Finite-temperature and bidirectional magneto-thermal closures

A broader class of regularized micromagnetic models addresses the failure of fixed-length ferromagnetic micromagnetics near and above the Curie point. The Landau–Lifshitz–Bernoulli model augments LL by a longitudinal relaxation term,

$\mathbb R^d$1

where $\mathbb R^d$2 is the equilibrium magnetization magnitude supplied by experiment, mean-field theory, or atomistic models. The paper’s central claim is that this variable-length continuum model recovers standard micromagnetics in the ferromagnetic regime and classic Maxwell magnetostatics of paramagnets at high temperature, thereby bridging atomic- and macro-scale magnetism across $\mathbb R^d$3 (Kiefe et al., 28 May 2026).

A related but thermodynamically sharper regularization appears in the LLBar-based finite-temperature ferromagnetic model. There, longitudinal relaxation, nonlocal exchange damping, and explicit electron–phonon temperature equations are combined so that magnetic dissipation is routed into the corresponding thermal reservoirs. The model is non-stochastic and restricted to the ferromagnetic state, but it is presented as thermodynamically self-consistent precisely because it resolves energy exchange between spin, electron, and phonon subsystems rather than treating damping as an abstract sink (Dvornik et al., 2014).

An even stronger thermodynamic closure is developed in the bidirectional magneto-thermal coupling model, which couples stochastic LLG to a generalized heat equation,

$\mathbb R^d$4

Here the local temperature is itself a dynamical field, the thermal noise amplitude scales with the instantaneous local $\mathbb R^d$5, and magnetic dissipation plus stochastic work feed back into the bath. Using Itô stochastic calculus, the paper proves first-law consistency and recovery of Boltzmann statistics at equilibrium; numerically it captures finite-bath temperature reduction and exchange-modified density of states (Yi et al., 11 Jun 2026).

These models regularize micromagnetics in a constitutive and thermodynamic sense. They do not primarily cure a PDE ill-posedness; instead, they close the finite-temperature problem by allowing longitudinal amplitude variation and by tracking the thermal reservoir as an active field rather than a one-way input.

6. Multiscale, atomistic, and multi-sublattice regularizations

Another response to the breakdown of standard micromagnetics is to substitute a more resolved description precisely where continuum assumptions fail. In elevated-temperature recording, an atomic-level model resolves each grain into a $\mathbb R^d$6 array of atomic spins with cell size $\mathbb R^d$7 nm, evolves them by stochastic LLG, and includes strong intra-grain exchange $\mathbb R^d$8 kOe. Relative to a uniformly magnetized grain with prescribed $\mathbb R^d$9 and $S^2$0, the atomistic model captures localized, non-uniform reversal, slightly reduces the field required to switch grains at elevated temperatures, and yields considerably more fluctuations; in the ultra-high-anisotropy case, the write bubble is about $S^2$1 larger (Mercer et al., 2012).

The multiscale HAMR framework pushes this logic further by partitioning a moving simulation window into a central atomistic region and flanking micromagnetic regions. In the main setup, a $S^2$2 nm window is split into a $S^2$3 nm atomistic center and two $S^2$4 nm micromagnetic side regions. The near- and above-$S^2$5 region is treated atomistically, the cooler region micromagnetically, and information is exchanged by coarse-graining and inverse reconstruction across the interface. The reported performance figure is concrete: a $S^2$6 nm-wide track of twenty-five $S^2$7 nm-long bits can be recorded in several hours on a single GPU (Gija et al., 4 Feb 2025).

Ferrimagnetic compensation supplies a different multiscale or structure-resolved regularization theme. Instead of collapsing the material into one effective ferrimagnetic medium, the two-sublattice model evolves two unit magnetization fields $S^2$8 with sublattice-specific $S^2$9, explicit inter-sublattice exchange $|m|=1$0, and optional sublattice-resolved torques. Its importance lies in avoiding the need for diverging effective parameters near angular momentum compensation: the separate LLG equations remain well defined even when effective one-medium descriptions become singular or misleading (Martínez et al., 2019).

A related coarse-grained regularization appears in nanocomposite permanent magnets. There, unresolved grain-boundary disorder is replaced by a core-shell micromagnetic model in which grain interiors keep bulk $|m|=1$1 parameters while the shell has reduced anisotropy and exchange, and shell-shell exchange is weakened. The authors explicitly interpret the shell as a surrogate for structurally and chemically deteriorated near-surface regions, and the best agreement with measured hysteresis loops requires this core-shell model plus an added phenomenological superparamagnetic contribution (Erokhin et al., 2017).

What unifies these otherwise different constructions is that the unresolved regime is not ignored. It is either resolved atomistically, represented by explicit sublattices, or homogenized into shells and constitutive interface laws.

7. Numerical stabilization, reduced-order models, and unresolved limits

Not every regularized micromagnetic model modifies the continuum PDE itself. Some works regularize the approximation or representation. The symmetric Gauss–Seidel projection method leaves the micromagnetic PDE unchanged but constructs a more stable discretization by combining implicit heat-flow solves, a two-pass symmetric Gauss–Seidel iteration, and projection onto the sphere. The method retains first-order temporal accuracy and second-order spatial consistency while providing improved discrete energy stability relative to standard GSPM, particularly in weak-damping or undamped tests where one-sided Gauss–Seidel can show nonphysical energy behavior (Miao et al., 26 Jun 2026).

Other works regularize the representable state space. MagTense discretizes the body into uniformly magnetized rectangular cuboid or tetrahedral tiles and computes the demagnetizing interaction analytically for that piecewise-constant ansatz. The paper is explicit that this is not an added regularization term, but it is an implicit coarse-graining because all sub-tile structure is suppressed and the demagnetizing field is exact only for the reduced representation (Bjørk et al., 2021). The normal-modes framework does something analogous in dynamics: it projects LLGS onto a finite number of eigenmodes of the linearized micromagnetic operator, yielding a nonlinear ODE system with linear, quadratic, cubic, quartic, and quintic modal couplings and allowing accuracy to be tuned between macrospin and full spatial discretization (Perna et al., 2021).

The limits of regularized micromagnetics are equally instructive. In frustrated-exchange skyrmion systems, the collapse barrier can depend on atomistic exchange components invisible to the continuum parameters $|m|=1$2. The paper on long-range frustration shows that skyrmions with identical micromagnetic parameters can exhibit significantly different energy barriers because the saddle-point texture is much more strongly noncollinear than the metastable skyrmion. A plausible implication is that no low-order local continuum regularization can be assumed complete for barrier physics unless it encodes the ultraviolet information relevant to the saddle point (Zhu et al., 30 Jun 2026).

The literature therefore supports a restrained conclusion. A regularized micromagnetic model is not a single equation but a design principle: identify the regime in which classical micromagnetics fails, then relax or replace exactly the assumption responsible for that failure. The resulting regularization may be measure-valued, higher-order, bounded-amplitude, thermodynamically closed, multiscale, or purely numerical. Its value is judged not by formal uniformity across all contexts, but by whether it preserves the physically relevant micromagnetic structure while removing the specific nonregularity that the unmodified model cannot accommodate.