Sharp Phase-Field Method (S-PFM)

• S-PFM is a family of phase-field techniques that retain diffuse-interface variables while rigorously recovering sharp-interface behavior, reducing grid pinning and numerical artifacts.
• It integrates intrinsically discrete formulations, variational asymptotic limits, and hybrid strategies—including neural and high-order solver approaches—to enhance simulation fidelity.
• The method enables simulations with nearly one-grid-point interface resolution, ensuring translational invariance and frictionless kinetics for accurate interfacial dynamics.

Sharp Phase-Field Method (S-PFM) denotes a family of phase-field formulations in which diffuse-interface variables are retained, but the model, discretization, or asymptotic analysis is constructed so that sharp-interface behavior is recovered much more faithfully than in conventional phase-field methods. In one prominent usage, S-PFM is an intrinsically discrete formulation: the free energy is built directly on the grid, discrete equilibrium interfaces are translationally invariant, and interfaces can be represented with essentially one grid point without grid pinning (Finel et al., 2018, Fleck et al., 2019). In another usage, the same label refers to diffuse-interface approximations that are proved, or shown by matched asymptotics, to converge to sharp-interface energies or free-boundary problems, including Eulerian interfacial energies in elastic solids, compressible Navier–Stokes/Allen–Cahn systems, and curvature-coupled biomembranes (Grandi et al., 2018, Yazhou et al., 2021, Elliott et al., 2020). More recent hybrid finite-element and neural approaches pursue the same objective by treating mechanics, interface geometry, or solver architecture in a semisharp or discontinuity-aware manner while retaining phase-field evolution (Dobrzanski et al., 2024, Salami et al., 2020, Bao et al., 2024, Lei et al., 28 Nov 2025).

1. Terminological scope and defining characteristics

The recent literature uses “Sharp Phase-Field Method” for several closely related constructions. What unifies them is not a single PDE, but a common objective: to preserve the topological flexibility and variational structure of phase-field models while removing, or rigorously controlling, the artificial smearing, grid pinning, and asymptotic inconsistency that accompany naive diffuse-interface discretizations.

Usage in the literature	Representative formulation	Defining sharpness mechanism
Intrinsically discrete S-PFM	Discrete free energy on a lattice with tailored non-polynomial wells	Exact translational invariance of discrete interfaces and essentially one-grid-point resolution
Variational or asymptotic sharp-interface approximation	Modica–Mortola-type diffuse interface converging to a perimeter or free-boundary model	$\Gamma$-convergence or matched asymptotics as $\varepsilon \to 0$
Hybrid semisharp or solver-side approach	LET-PF, preconditioned conservative phase field, discontinuity-aware neural solvers	Sharp mechanics, sharper interface geometry, or reduced effective interface thickening

A recurrent misconception is to equate “sharp” with a mathematically zero-thickness interface at the PDE level. That is not the general meaning. In the discrete SPFM of Finel and coauthors, the interface remains a diffuse tanh-like profile, but the grid representation becomes translationally invariant and can be reduced to one grid spacing without pinning (Finel et al., 2018). In rigorous variational settings, “sharp” instead refers to the limiting object selected as the diffuse thickness parameter tends to zero, such as an Eulerian perimeter or a free boundary with Young–Laplace jump conditions (Grandi et al., 2018, Yazhou et al., 2021).

2. Intrinsically discrete formulations

The discrete S-PFM introduced in early 2018 departs from the conventional strategy of first writing a continuum functional and then discretizing it. Instead, the discrete functional is primary. In one dimension, the free energy is posed as

$$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$

with $\phi_n=\phi(nd)$ and discrete gradient $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$. The design requirement is that the discrete equilibrium interface be exactly

$\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$

for any real shift $x_0$. This leads to a non-polynomial double-well potential with logarithmic terms,

$$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$

so that the discrete tanh profile is an exact equilibrium and the interface energy is independent of its position relative to the grid (Finel et al., 2018).

This construction changes the numerical meaning of interface width. In conventional diffuse-interface PFM, the interface width must be resolved by many points; the reported rule of thumb is $6$–$8$ points across the interface to avoid pinning and anisotropic artifacts. In S-PFM, the transition can occur over essentially one grid point. For $\varepsilon \to 0$0, the discrete profile drops from near $\varepsilon \to 0$1 to near $\varepsilon \to 0$2 between neighboring nodes, yet remains translationally invariant and free of preferred grid positions (Finel et al., 2018).

The three-dimensional extension is formulated on a face-centred cubic computational grid. The discrete free energy uses first-, second-, and third-neighbour shells,

$\varepsilon \to 0$3

with $\varepsilon \to 0$4. Exact translational invariance cannot be imposed for all orientations simultaneously, so one family of lattice planes is treated exactly and the weights $\varepsilon \to 0$5 are optimized so that the average interface energies of three selected plane families coincide. The paper reports optimized weights for the $\varepsilon \to 0$6- and $\varepsilon \to 0$7-based constructions at $\varepsilon \to 0$8 and shows nearly isotropic interfacial energies across the chosen families (Finel et al., 2018).

Operationally, the method reproduces interfacial kinetics with unusually coarse interface resolution. For a one-dimensional interface, the interface energy $\varepsilon \to 0$9 is invariant to numerical precision, reported to $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$0 digits under shifts of $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$1. For a curvature-driven shrinking precipitate, S-PFM reproduces the exact continuum limit for the precipitate volume fraction versus time; for $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$2 and $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$3, the results are indistinguishable from the continuum limit. Classical PFM requires approximately $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$4 grid points across the interface to achieve the same accuracy, around $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$5 error in interface velocity, whereas S-PFM uses one grid point. In three dimensions, sphericity loss stays below $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$6, and reaches about $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$7 for $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$8; the corresponding coarse-grid capability implies roughly $$F = d \sum_n \left[ g(\phi_n) + \frac{\lambda}{2 d^2}\, \|\tilde{\nabla} \phi_n\|^2 \right],$$9 fewer grid points for a fixed physical problem in three dimensions (Finel et al., 2018).

3. Translational invariance, grid friction, and frictionless kinetics

A decisive refinement of the discrete S-PFM program is the explicit treatment of spurious grid friction. On a discrete grid, conventional finite-difference or FFT implementations break translational invariance: as a diffuse interface moves across grid points, the discrete free energy oscillates with interface position, generating numerical friction, pinning, and velocity oscillations. The 2019 analysis of frictionless motion identifies this broken translational invariance as the origin of the problem and reconstructs the discrete free-energy density so that the analytic profile remains an exact discrete solution (Fleck et al., 2019).

The reference moving interface is

$\phi_n=\phi(nd)$0

with interface normal $\phi_n=\phi(nd)$1 and center position $\phi_n=\phi(nd)$2. The key grid-coupling parameter is

$\phi_n=\phi(nd)$3

where $\phi_n=\phi(nd)$4 is a grid vector. Using the tanh addition property, neighboring values can be written exactly in terms of $\phi_n=\phi(nd)$5, and this yields a direction-dependent equilibrium potential

$\phi_n=\phi(nd)$6

With this correction, the analytic tanh profile satisfies the discrete force-equilibrium equation exactly for the prescribed orientation (Fleck et al., 2019).

Two constructions are distinguished. In the TI$\phi_n=\phi(nd)$7 models, translational invariance is restored for planar interfaces aligned with selected crystallographic directions. In the TI$\phi_n=\phi(nd)$8 model, the local interface normal is estimated in the interface region and the coupling parameters are updated accordingly, producing local translational-invariance restoration for arbitrarily oriented planar interfaces. Under homogeneous driving forces, this leads to frictionless motion of arbitrarily oriented diffuse interfaces on a fixed $\phi_n=\phi(nd)$9D grid, together with superior isotropy of both interface energy and kinetics (Fleck et al., 2019).

This framework also clarifies the meaning of “sharpness” in the discrete literature. The method does not make the interface mathematically discontinuous; rather, it permits extremely low numerical resolution, down to $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$0, without pinning, and it eliminates the spurious dependence of velocity and energy on the interface’s position relative to the mesh. In that sense, S-PFM is a discrete translational-invariance restoration procedure as much as it is an interface-thinning strategy (Fleck et al., 2019).

4. Variational and asymptotic sharp-interface limits

A distinct S-PFM lineage treats the phase field as a regularization of a sharp-interface variational problem and asks for existence, compactness, and the correct limiting interfacial energy as $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$1. In the Eulerian interfacial-energy setting of two-phase elasticity, the reference body $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$2 is deformed by $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$3 into $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$4, and the sharp phase indicator is Eulerian, $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$5. The sharp-interface energy is

$\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$6

so the bulk term is Lagrangian while the interfacial term is purely Eulerian. The diffuse approximation replaces $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$7 by $\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$8 on the deformed configuration and uses

$\tilde{\nabla}\phi_n=\phi_n-\phi_{n-1}$9

Under polyconvexity, coercivity, frame indifference, blow-up as $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$0, the Ciarlet–Nečas condition, and finite distortion, admissible deformations form a class of homeomorphisms. The diffuse energies admit minimizers, and $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$1-converge to the sharp Eulerian perimeter functional; minimizers of $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$2 converge to minimizers of $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$3 as $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$4 (Grandi et al., 2018).

Matched-asymptotic sharp-interface limits play the same role in fluid models. For the compressible non-isentropic Navier–Stokes/Allen–Cahn system with diffuse-interface thickness $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$5, the limiting bulk equations are the compressible Navier–Stokes equations with heat conduction in each phase, while the interface becomes a free boundary. Velocity and temperature are continuous across the interface, and the mechanical jump condition is the Young–Laplace law,

$\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$6

The kinematic condition depends on the mobility scaling. For $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$7, the interface is materially advected: $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$8 For $\phi_n = \frac{1+\tanh((nd-x_0)/w)}{2}$9, a phase-transition regime appears on the subset where density is continuous,

$x_0$0

whereas on the complement, where density jumps, one again has $x_0$1 together with $x_0$2 in the leading-order limit (Yazhou et al., 2021).

An analogous diffuse-to-sharp program exists on curved surfaces. For near-spherical two-phase biomembranes, a phase field $x_0$3 is coupled to a height field $x_0$4 over a sphere, and the diffuse interfacial part is of Modica–Mortola type,

$x_0$5

Mechanical equilibration allows elimination of $x_0$6 in favor of a reduced diffuse energy $x_0$7, and the $x_0$8-limit is a sharp-interface energy equal to line tension $x_0$9, with $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$0, plus a nonlocal elastic term. Formal asymptotics of conserved Allen–Cahn dynamics then produce a sharp-interface gradient flow coupling geodesic curvature flow of the phase boundary to a fourth-order PDE free-boundary problem for the membrane deformation (Elliott et al., 2020).

5. Hybrid semisharp discretizations and high-order interface tracking

A third strand seeks sharpness through hybridization rather than through purely diffuse energetics or purely discrete translational invariance. In the LET-PF approach for microstructure evolution, the phase-field variable still obeys a Ginzburg–Landau evolution law and retains the usual diffuse interfacial energy,

$$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$1

but mechanically it acts as a level-set-like function: the contour $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$2 defines a sharp interface. Elements cut by that interface are treated as simple laminates with phase volume fraction $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$3 and lamination orientation $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$4, and only a one-element-thick layer around the interface mixes phases. This “diffuse–semisharp” construction yields higher accuracy than conventional PFM for the problems studied and gives qualitatively correct results on significantly coarser meshes and at lower computational cost (Dobrzanski et al., 2024).

High-order interface-tracking methods have adopted a related philosophy. The preconditioned conservative phase-field method developed in a Flux Reconstruction framework writes the transport equation in conservative form and sharpens the interface through a balance of diffusion and antidiffusion,

$$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$5

with normals obtained from a preconditioned signed-distance-like field $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$6 and localized artificial viscosity used only in the reinitialization step. The reported parameter choice is $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$7 and $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$8. Implemented in PyFR, the method remains conservative, captures highly distorted interfaces with superior accuracy compared with conventional and high-order VOF and level set methods, and is designed for high-order unstructured grids and massively parallel computation (Salami et al., 2020).

Phase-field fracture at large deformation exhibits an analogous search for effective sharpness. In adaptive ES-FEM, the crack phase field $$g(\phi) = \frac{\lambda}{4 d^2} \left\{ \frac{\alpha^2 - 1}{\alpha^2} \log\left[1 - \alpha^2 (2\phi -1)^2\right] - (2\phi -1)^2 \right\},\qquad \alpha=\tanh(d/w),$$9 is coupled to a Griffith-type large-strain formulation with degradation

$6$0

and local refinement is triggered where the phase field indicates a developing crack. Although the method remains diffuse, the combination of edge-based strain smoothing and targeted mesh refinement yields a narrow, well-defined damage band and about a $6$1 computational gain over non-adaptive ES-FEM in the reported benchmarks, which is conceptually aligned with S-PFM objectives (Tian et al., 2019).

6. Neural and data-driven sharp-interface-capable formulations

Recent machine-learning work shifts the emphasis from model-side sharpness to solver-side sharpness. Phase-Field Weak-form Neural Networks parameterize weak solutions of Allen–Cahn and Cahn–Hilliard equations by deep neural networks equipped with a periodic layer and compactly supported local test functions. The free energy remains classical,

$6$2

but the weak-form loss, local training, and periodic Fourier-like features are designed for the strongly nonlinear and higher-order character of the equations. The stated advantage is that the framework can efficiently solve phase-field equations characterizing sharp transitions and can identify important parameters in forward and inverse settings (Bao et al., 2024).

The discontinuity-aware PINN developed for three-phase flows pursues the same goal more explicitly. It solves an energy-stable AC–CH–NS model with phase change, uses Fourier embeddings, discontinuity-aware gated residual blocks, a learnable local artificial-viscosity term, adaptive time marching, and loss balancing, and is intended to mitigate spectral bias and automatic interface thickening. The artificial viscosity is localized to interfacial regions through an interface sensor, and the method is presented as a way to make classical phase-field PDEs accessible in the sharp-interface regime rather than as a new asymptotic phase-field model. In the reported tests, conventional PINNs fail as the interface thickness is reduced, whereas the discontinuity-aware solver resolves sharp interfacial dynamics and extends to a three-phase droplet-icing case with viscosity and density ratios exceeding $6$3 and $6$4 orders of magnitude, capturing the pointy-tip morphology (Lei et al., 28 Nov 2025).

Across these literatures, “sharp” therefore names a target property rather than a single formalism. It can mean exact discrete translational invariance, rigorous convergence to a perimeter or free-boundary model, semisharp mechanics on nonconforming meshes, or solver architectures that avoid magnifying the nominal interface thickness. The limitations are correspondingly diverse: full rotational invariance cannot be made exact on a discrete grid in $6$5D, exact frictionless motion is proved most cleanly for planar interfaces under homogeneous driving forces, and semisharp finite-element formulations still face nontrivial issues for multiple interfaces per element and triple junctions (Finel et al., 2018, Fleck et al., 2019, Dobrzanski et al., 2024).