Landau–Brazovskii Free Energy

• The Landau–Brazovskii free energy model is a continuum free-energy functional that describes systems developing spatially modulated order at nonzero wavevectors.
• It combines higher-order gradient operators with quadratic, cubic, and quartic bulk terms to capture complex phase transitions and nucleation dynamics.
• Advanced numerical techniques, including Fourier pseudo-spectral methods and adaptive optimization schemes, efficiently resolve its intricate phase diagram and saddle point computations.

The Landau–Brazovskii free energy is a continuum free-energy functional for systems that develop spatially modulated order at a nonzero wavevector. In its standard form it combines a higher-order gradient operator, which selects a preferred modulation length scale, with quadratic, cubic, and quartic bulk terms, and it is typically supplemented by a zero-mean constraint on the order parameter. This framework is used to describe phase behavior and structural transitions in microphase-separating and pattern-forming systems, including block copolymers, magnetic systems, liquid crystals, quasicrystals, and “nuclear pasta” in neutron stars (Zhang et al., 20 Oct 2025).

1. Canonical formulation

A dimensional form of the functional is

$$\mathcal{E}[\phi] = \int \mathrm{d}\mathbf{r}\left\{ \frac{\xi_0^2}{8q_0^2}\left|(\nabla^2+q_0^2)\phi\right|^2 +\frac{\tau}{2}\phi^2 -\frac{\gamma}{3!}\phi^3 +\frac{\lambda}{4!}\phi^4 \right\},$$

where $\phi(\mathbf r)$ is a scalar order parameter, $\xi_0$ is the bare correlation length, $q_0$ is the critical wavenumber, $\tau$ is a reduced temperature parameter, $\gamma$ quantifies asymmetry through the cubic term, and $\lambda>0$ stabilizes the energy through the quartic term. After non-dimensionalization and rescaling, a widely used form is

$$\mathcal{E}[\phi] = \int \mathrm{d}\mathbf{r}\left\{ \frac{1}{2}\left|(\nabla^2+1)\phi\right|^2 +\frac{\tau}{2}\phi^2 -\frac{\gamma}{3!}\phi^3 +\frac{1}{4!}\phi^4 \right\}, \qquad \int \phi = 0,$$

with the integral constraint enforcing mass conservation (Zhang et al., 20 Oct 2025).

A common numerical formulation writes the same structural content as

$$E(\phi)=\int_{\Omega}\left\{ \frac12(\Delta\phi+\phi)^2 -\frac{\alpha}{2}\phi^2 +\frac{1}{4!}\phi^4 -\frac{\gamma}{3!}\phi^3 \right\}\,d\mathbf x,$$

again with zero average $$\bar\phi=\frac1{|\Omega|}\int_\Omega\phi\,d\mathbf x=0$$ (Zhang et al., 2022). On curved manifolds, the spherical Landau–Brazovskii model replaces the Euclidean Laplacian by the spherical Laplacian and considers

$\phi(\mathbf r)$0

subject to zero mean on $\phi(\mathbf r)$1 (Qiu et al., 25 Jan 2025).

The physical interpretation of the terms is standard within the model class. The higher-order operator $\phi(\mathbf r)$2 or $\phi(\mathbf r)$3 captures competition between short-range interactions favoring uniformity and longer-range interactions favoring modulations at a preferred wavelength. The quadratic term governs the tendency toward order or disorder, the cubic term breaks the symmetry $\phi(\mathbf r)$4 and enables first-order transitions, and the quartic term stabilizes the free energy at large amplitude (Zhang et al., 20 Oct 2025).

2. Ordered phases and phase selection

The defining feature of the Landau–Brazovskii model is that it stabilizes periodic structures at a nonzero wavevector, rather than only homogeneous states or weakly modulated states. This is the basis for its use in microphase separation and related modulated-phase systems (Zhang et al., 20 Oct 2025).

For the three-dimensional model, a computed phase diagram was reported to encompass eight distinct stable phases: the disordered phase, lamellar (LAM), hexagonally-packed cylinders (HEX), body-centered cubic (BCC), face-centered cubic (FCC), double gyroid (DG), Frank–Kasper A15, and Frank–Kasper $\phi(\mathbf r)$5 (Zhang et al., 20 Oct 2025). In that construction, the stable phase at fixed $\phi(\mathbf r)$6 is the one with the lowest free energy among candidate structures, and the domain size for each periodic phase is optimized because the free energy depends on the fit between the computational domain and the phase periodicity (Zhang et al., 20 Oct 2025).

Reported phase	Context	Source
Disordered	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
LAM	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
HEX	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
BCC	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
FCC	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
DG	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
A15	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
Frank–Kasper $\phi(\mathbf r)$7	Stable phase in 3D phase diagram	(Zhang et al., 20 Oct 2025)
FDDD	Stable region identified later	(Bao et al., 4 Mar 2026)

Subsequent second-order optimization work enlarged the computationally accessible phase repertoire by identifying a stable region of the cubic FDDD phase, described there as previously unreported in the Landau–Brazovskii phase diagram (Bao et al., 4 Mar 2026). A plausible implication is that the known phase diagram depends not only on the free-energy functional itself but also on whether the numerical method is able to distinguish physically relevant local minima from saddle-type stationary points.

3. Free-energy landscape, saddles, and nucleation pathways

Within the Landau–Brazovskii framework, ordered and metastable phases correspond to local minima of a nonconvex free-energy landscape separated by barriers. Transition pathways are minimum energy paths connecting two minima through at least one index-1 saddle, and the saddle serves as the transition state (Zhang et al., 20 Oct 2025).

The transition state is identified with a critical nucleus: a localized region of the stable phase embedded in a metastable background. Along computed transition pathways, the critical nucleus can be analyzed in terms of shape, energy barrier, and Hessian eigenvalues. High-index saddles also occur and were associated with multi-nucleus configurations or collective transformations (Zhang et al., 20 Oct 2025). Hessian eigenvectors at the transition state determine directions of nucleus growth, shrinkage, and translation (Zhang et al., 20 Oct 2025).

Parameter dependence is central to the nucleation picture. When $\phi(\mathbf r)$8 approaches a phase boundary, the free-energy difference between competing phases decreases, critical nuclei become larger, and the nucleation barrier increases (Zhang et al., 20 Oct 2025). The reported examples include spherical nuclei in BCC–HEX transitions and ellipsoidal nuclei in HEX–LAM transitions, reflecting the anisotropy imposed by the surrounding ordered phase (Zhang et al., 20 Oct 2025). The energy barrier is the free-energy difference between the transition state and the initial metastable phase, and it governs the nucleation rate in the usual activated-process interpretation (Zhang et al., 20 Oct 2025).

This landscape perspective also clarifies a common oversimplification: the Landau–Brazovskii model is not only a phase-diagram generator but also a framework for computing metastability limits, transition states, and nucleation mechanisms between modulated phases (Zhang et al., 20 Oct 2025).

4. Gradient-flow dynamics and metric structure

For the computation of stationary states and dissipative evolution, one frequently considers a mass-conserving Allen–Cahn-type gradient flow,

$\phi(\mathbf r)$9

where $\xi_0$0 is a Lagrange multiplier chosen so that the spatial average of $\xi_0$1 is preserved (Zhang et al., 2022). In that formulation, the continuous dynamics satisfy both mass conservation and energy dissipation, with

$\xi_0$2

(Zhang et al., 2022).

For saddle-point computation, however, the choice of metric is not a technical detail. It was emphasized that, for the same energy functional, the saddle points, as well as other stationary points, are different in different metrics such as the $\xi_0$3 metric and the $\xi_0$4 metric (Gu et al., 2020). This is especially important in conserved systems, where the natural gradient flow is in the $\xi_0$5 metric.

A projection strategy addresses this issue by enforcing the mean-zero constraint in the $\xi_0$6 setting through

$\xi_0$7

Projected gentlest ascent dynamics and projected iterative minimization then search for saddle points in the mass-conserving subspace without explicit $\xi_0$8 inner products or repeated Poisson solves (Gu et al., 2020). In the reported two-dimensional Landau–Brazovskii computations, this produced the same saddle points as direct $\xi_0$9 approaches while reducing CPU time by about $q_0$0 (Gu et al., 2020). A recurring misconception is that saddle points are intrinsic to the energy alone; in this setting, the physically relevant transition states depend on the metric associated with the conserved dynamics.

5. Numerical discretization and stationary-state algorithms

Efficient simulation of the Landau–Brazovskii model is challenging because the functional contains cubic and quartic nonlinearities and a high-order spatial operator, making the dynamics stiff (Zhang et al., 2022). A standard spatial discretization is the Fourier pseudo-spectral method, in which differential operators are diagonal or simple multiplications in reciprocal space and nonlinear terms are evaluated in real space using FFT and inverse FFT (Zhang et al., 2022).

For time integration, a convex-splitting formulation decomposes the energy into convex and expansive parts and uses an implicit treatment of the linear convex contribution with explicit treatment of the nonlinear expansive contribution. In the reported scheme, this yields a linear system at each step together with unconditional energy stability and exact mass conservation (Zhang et al., 2022). The spectral deferred correction (SDC) method and its adaptive variant were then combined with convex splitting to obtain higher-order temporal accuracy while preserving energy stability; numerical experiments on two- and three-dimensional periodic crystals were reported to show the efficiency of the method (Zhang et al., 2022).

A distinct line of work treats stationary-state computation as a finite-dimensional nonconvex optimization problem after pseudospectral discretization. The implicit-explicit trust region method exploits a Hessian structure in which the interaction term is diagonal in reciprocal space whereas the bulk energy is diagonal in physical space, enabling efficient matrix-vector products via FFT with $q_0$1 cost per iteration (Bao et al., 4 Mar 2026). Its stated objective is convergence to second-order stationary points, namely local minima with positive semidefinite Hessian, rather than merely first-order stationary points that may be saddles (Bao et al., 4 Mar 2026). In numerical experiments on the Landau–Brazovskii model, the method was reported to escape saddle points and significantly outperform existing first-order schemes (Bao et al., 4 Mar 2026).

The distinction between first-order and second-order stationary points is physically consequential. In the Landau–Brazovskii setting, only second-order stationary points correspond to stable or metastable phases, whereas first-order stationary points may be transition states or other unstable configurations (Bao et al., 4 Mar 2026).

6. Spherical formulations, geometry adaptation, and phase discovery

The spherical Landau–Brazovskii model replaces periodic Euclidean coordinates by a field on $q_0$2 and discretizes the order parameter by spherical harmonic expansion,

$q_0$3

truncated at finite degree $q_0$4 (Qiu et al., 25 Jan 2025). Because the spherical Laplacian acts diagonally on $q_0$5, the quadratic part of the discretized energy is diagonal in harmonic space, while the cubic and quartic terms couple modes through Wigner-$q_0$6-type coefficients (Qiu et al., 25 Jan 2025). To compute stationary states, five optimization methods were developed: accelerated adaptive Bregman proximal gradient, Nesterov, adaptive Nesterov, adaptive nonlinear conjugate gradient, and adaptive gradient descent (Qiu et al., 25 Jan 2025).

A notable analytical device in this setting is principal mode analysis (PMA), which links the optimal sphere radius $q_0$7 to the dominant harmonic degree $q_0$8 through

$q_0$9

The PMA method was proposed both to estimate good initial configurations and to reveal the relationship between the optimal sphere radius and the dominant degree of spherical harmonics (Qiu et al., 25 Jan 2025). This provides a direct mechanism for symmetry selection on curved geometry.

On periodic Euclidean domains, a separate difficulty is domain incompatibility. Fixed computational domains may impose artificial stress and trap the system in high-energy metastable states. A geometry-adaptive deep variational framework addresses this by jointly optimizing the order parameter and the domain side lengths $\tau$0, treating geometry as trainable variables in the variational problem (Xie et al., 5 Mar 2026). In that framework the field is represented by a neural network with hard imposition of the zero-mean constraint, the energy is evaluated by FFT-based pseudospectral methods, and a warmup penalty is introduced to destabilize the disordered phase under small initializations (Xie et al., 5 Mar 2026). Guided initialization is then used for topologically intricate phases with narrow basins of attraction (Xie et al., 5 Mar 2026).

This geometry-adaptive approach was reported to discover, in three dimensions, LAM, HEX, BCC, FCC, A15, and, with guided initialization, DG and Frank–Kasper $\tau$1, while automatically adapting the computational domain to match the natural periodicity of the solution (Xie et al., 5 Mar 2026). A practical misconception addressed by this line of work is that the periodic box is merely a numerical convenience; the reported results argue that domain size is part of the variational problem because mismatch induces artificial elastic stress (Xie et al., 5 Mar 2026).

7. Generalized theoretical setting and interpretation of coefficients

The Landau–Brazovskii free energy can be placed within a broader theory of nonlocal free-energy functionals written in terms of convolution kernels rather than assumed gradient expansions. In that framework, Ginzburg–Landau, Swift–Hohenberg, and Landau–Brazovskii forms emerge when the generalized functional is reduced to local gradient-based representations under appropriate assumptions on locality and kernel smoothness (Li et al., 2022).

For isotropic systems truncated at fourth order in derivatives, the resulting free energy takes the form

$\tau$2

which includes the characteristic higher-derivative structure associated with Landau–Brazovskii and Swift–Hohenberg descriptions (Li et al., 2022). In this interpretation, the gradient and higher-gradient coefficients are not purely phenomenological: the second-order coefficient is associated with second moments of the interaction kernel, while the coefficient of $\tau$3 is associated with fourth moments (Li et al., 2022).

This suggests a microscopic reading of the Landau–Brazovskii parameters: the preference for a finite modulation length and the stiffness against deviations from that length scale can be connected to the spatial structure of underlying interactions rather than introduced only as formal continuum parameters. Within that generalized theory, the Landau–Brazovskii model is therefore both a phenomenological phase-field functional and a controlled truncation of a broader nonlocal energetic description (Li et al., 2022).